GC-INF: A Novel Adaptive Traffic Control System using Machine Learning for Turning Ratio Predictions

The long-standing problem of traffic congestion has major environmental and financial implications globally, and traditional traffic control methods are becoming outdated and inefficient. Recently, advanced urban infrastructure development has allowed for Intelligent Traffic Systems (ITSs) to be introduced as an alternative to manage heavy traffic flow and ease congestion. This paper proposes a novel traffic control system, GC-INF, that uses machine learning to make turning ratio predictions at intersections. The GC-INF model is based on the integration of two cutting-edge models, the Graph Convolutional Network (GCN) and the Informer model, which enable it to capture both spatial and temporal dependencies in the input data while avoiding information loss. The model out-performs four state-of-the-art baseline models in all evaluation criteria, including RMSE, MAE, R2, var, and Accuracy, with minimal accuracy loss over short and long-term horizons. This highly optimal solution for predicting turning ratios at detection points in a road network demonstrates that models that can capture both spatial and temporal dependencies have stronger prediction capabilities compared to models that can only capture temporal dependencies. In future research, the GC-INF model could be validated using real-world data collected from loop detectors at intersections, and ensemble techniques like bagging and boosting could be applied to enhance the performance, interpretability, and uncertainty quantification of the model for practical applications. The results show the potential of machine learning models in improving traffic control and management in urban areas, which can lead to increased travel efficiency and reduced environmental and financial costs.