Using CNN-LSTM to predict signal phasing and timing aided by High-Resolution detector data

Abstract

This paper proposes a real-time signal timing prediction based on deep learning algorithms that takes various traffic flow parameters as input and predicts signal timing parameters. Signal retiming methods have been traditionally used to improve traffic flow at intersections by reducing delay and improving the level of service at signalized intersections. This can often be a lengthy process that includes several iterations of various optimization methods. In this paper, we have used detector data to calculate the traffic flow metrics at several intersections. This processed data is then used to predict the signal timing and phasing for the next six cycles with reasonable accuracy. Seventeen intersections from two distinct corridors have been used in this study. One of the corridors runs adaptive signal control and the other corridor runs actuated signal control. The proposed CNN-LSTM model shows that cycle length can be accurately predicted with an MAE of 7 to 16 s and phase duration can be predicted with an MAE of 3 to 8 s. The trained model was also successfully validated at five different unknown intersections.