Real-Time Prediction of the Lane-Based Delay for Group-Based Adaptive Traffic Operations Using Long Short-Term Memory

Abstract

This study proposes a deterministic real-time lane-based control delay model for traffic operations based on Long Short-Term Memory (LSTM). Our proposed framework includes a model-based approach to compute the control delay in an individual lane for a single cycle and a data-driven approach to predict the queueing profiles and adjustment factors used in the future control delay formula. This framework not only secures an excellent performance of the proposed model under a wide range of data availability but also guarantees a lower computational burden for a real-time non-linear optimisation process in adaptive control logic. The modified deep learning method has three primary components in the proposed architecture of the lane-based control delay model cycle-by-cycle. First, the data-driven and model-based approaches are integrated to improve the reliability and the accuracy of the control delay predictive formula. Second, the novel LSTM network is constructed to predict a cycle-based control delay in an individual lane while minimising inherent errors in the algorithm. Third, the predicted queue lengths at inflection points and adjustment factors are used to construct the delay polygons in the future cycle. Numerical simulations are set up using both synthetic and real-world data to give insights into the proposed model's performance compared to the existing models.