Short-term Traffic Volume Prediction at a Signalised Intersection Using LSTM

Abstract

Control by traffic signals is an important factor for short-term traffic state prediction at an intersection. Deep learning, which has the capability to capture nonlinear relationships, is achieving superior performance for an extensive range of prediction problems. Consequently, its application to traffic state prediction is expected. This study proposes a deep learning model that predicts traffic volume of the target road link downstream of an intersection, using the time-series traffic volume observation of the target link and multiple upstream links, and traffic signal control information at the intersection. The proposed model was validated with traffic volume data obtained from a traffic flow simulator. The validation results showed that the traffic volume data from upstream links and traffic signal control information contributed to improvements in prediction accuracy. However, when the signal control patterns of validation data were different from those of training data, the traffic signal control information made prediction accuracy worse.