Real-time signal queue length prediction using long short-term memory neural network

Optimal traffic control and signal planning can significantly reduce traffic congestion and potential delays at intersections. However, a major challenge to optimize traffic signal timing is to accurately predict traffic before commencing the next cycle. An optimal strategy cannot be achieved with a poor prediction of future traffic. In this study, using a deep learning approach, we develop a data-driven real-time queue length prediction technique. We consider a connected corridor where information from vehicle detectors (located at the intersection) will be shared to consecutive intersections. We assume that the queue length of an intersection in the next cycle will depend on the queue length of the target and two upstream intersections in the current cycle. We use InSync adaptive traffic control system data to train a long short-term memory neural network model capturing time-dependent patterns of a queue of a signal. To avoid overfitting and select the best combination of hyperparameters, we use sequential model-based optimization technique. Our experiment results show that the proposed model performs very well to predict the queue length. Although we run our experiments predicting the queue length for a single movement, the proposed method can be applied for other movements as well.