Imminent Throughput (ITP), the number of vehicles in a movement that can pass through an intersection given a unit of effective green time, serves as a crucial control input in many real-time optimization-based control schemes. The accuracy of ITP prediction can significantly influence control performance. If a movement with a green signal cannot discharge as many vehicles as predicted by the controller, its performance may be significantly reduced. However, most existing studies have focused on the design of control schemes while neglecting the importance of precise ITP prediction. These studies either assume that ITP can be accurately predicted or use traditional indices (e.g., saturation flow rate) or heuristic methods to predict ITP, resulting in relatively low accuracy. This paper proposes the use of a Deep Neural Network (DNN) to predict ITP and demonstrates that the DNN with Multiple Classifications (NN-C) models can predict ITP with higher accuracy, lower mean absolute error, and lower root mean squared error than other prediction methods (regression, decision tree, and heuristic methods). Experiments also show that control performance can be improved with more accurate ITP predictions using the NN-C.
Learning Imminent Throughput for Real-time Intersection Control with Deep Neural Network
2023-09-24
1252143 byte
Conference paper
Electronic Resource
English
Drivers' Collision Avoidance Pattern Before Imminent Intersection Accidents
British Library Conference Proceedings | 2019
|Drivers’ Collision Avoidance Pattern Before Imminent Intersection Accidents
Springer Verlag | 2018
|