A New Approach to Traffic Flow Prediction Based on Periodic Queue Length: Prediction Effect and Application Research

To address the limitations of short-term traffic flow prediction for traffic cycle signal control, we propose a periodic traffic volume prediction method based on queue length. Using a stacked bidirectional LSTM neural network, we train the model on the queue data from the first n historical cycles to predict the queue length of the next m cycles. By fitting the cycle queue length and traffic flow, the correlation between them is obtained, and this is used to predict the traffic flow in the next m cycles. Simulation analysis with intersection data reveals: (1) Prediction error varies with traffic flow characteristics, being highest during low-volume periods; (2) Prediction error decreases initially and then increases with n, minimizing at n=4 with an average error of 7.3% for the first 5 cycles; (3) Error increases with m, significantly so when m > 5. This method achieves high-accuracy periodic traffic volume prediction, meeting real-time signal control needs at intersections.