Dual-Layer Machine Learning Approach for Predicting Switching Times of Vehicle-Actuated Signal Phases

To enable green light optimal speed advisory systems and similar cooperative traffic applications, accurate predictions of traffic signal switching times are mandatory. However, in the case of traffic-actuated traffic signals, which adjust their switching behavior based on the current traffic situation, estimations cannot be derived based on bare statistics. Those types of traffic signals are most common in urban areas. While recent studies have applied machine learning models to this task, vehicle-actuated signal phases remain difficult to forecast because of their irregular and request-driven nature. To address this challenge, this study proposes a dual-layer machine learning (DLML) prediction procedure that separates event detection and timing estimation into two consecutive stages. The first layer classifies whether a signal phase will be released within the next 30 s; only in those cases does the second layer predict the exact switching time. As a benchmark, we apply a conventional time-to-signal-change prediction procedure. Both methods are implemented using the XGBoost algorithm and evaluated on real-world data from two traffic-actuated signalized intersections. The DLML approach consistently outperforms the conventional approach across all classification and regression metrics and reduces unnecessary computation by focusing prediction efforts on essential cases.