Predicting the Frequency and Duration of Turn-Bay Spillovers and Mutual Lane Blockages over Congested Arterials for Proactive Coordinated Signal Controls

Accurate prediction of queue sizes, spillovers, and lane blockages is critical for urban signal control and traffic management but remains challenging due to the time-varying and spatially interdependent nature of traffic dynamics. These challenges are especially pronounced in congested commuting arterials, where upstream disruptions rapidly propagate and affect traffic states and queue formation at multiple downstream intersections. To contend with such challenges in design of proactive real-time traffic control, this study presents a multibranch, multihead long-short-term memory (LSTM) system for queue dynamic prediction, including queue distance, onset time, and duration of bay spillovers or lane blockages. By assigning each upstream intersection its own dedicated LSTM branch, the proposed prediction model can learn and preserve location-specific temporal patterns at the target location and then exert its multihead fusion layer to receive and integrate these features to capture the information of cumulative traffic states and their impacts from upstream intersections. The proposed LSTM system is structured to reflect the spatial relations between a target intersection and its upstream intersections, and to maintain a balance between spatial resolution and computational efficiency. Performance evaluations using extensive simulations over MD 355 in Bethesda, Maryland, show that the proposed system achieves a 100% detection rate for all queue blockages, with exact predictions of onset times and durations and no false alarms. Compared with benchmark models such as the extended Kalman filter and standard recurrent neural networks, the proposed system demonstrates strong robustness in capturing spatial dependencies and persistent congestion patterns, key attributes for proactive real-time traffic signal control.