Predicting the Future Signalization of Traffic-Actuated Signals Using Extreme Gradient Boosting

Stops, braking, and acceleration maneuvers at traffic signals cause high fuel consumption and emission levels. Green light optimized speed advisory (GLOSA) can help to reduce a proportion of such unnecessary vehicle maneuvers. For this, GLOSA systems require reliable switching time estimations as input. However, estimating the switching times of traffic-actuated signals can be challenging. The signalization of traffic-actuated lights is adjusted to match the current traffic. Based on green-time requests, the status of signals can change with almost no lead time. This paper presents an approach for predicting the signalization of traffic-actuated signals. We exploit the limited nature of the space for switched signal state combinations at intersections, and express combinations of motorized traffic-related signal states as one feature to depict the motorized traffic signalization state (MTSSt) at an intersection. Predicting MTSSt-switches allows later determination of the switching times of individual signals. To conduct the predictions, the machine learning method “extreme gradient boosting” was used. A three-step methodology—data preparation, tuning the prediction procedure, and testing the approach—was applied and evaluated on the historical data of four traffic-actuated signalized intersections. The results showed that within a period of the next 30 s, signal changes were predicted with an overall precision and sensitivity of about 95% and an average mean absolute error of less than 1.1 s. In this period, the sequence of switched signals was predicted accurately to the second, without any deviation, in, on average, over 82% of cases.