Traffic Prediction and Optimization at Signalized Intersections with Connected Vehicles A Hybrid Approach

Abstract

With the evolution of Connected Vehicle (CV) technology, the exploration of CV-enabled traffic control systems emerges as a novel strategy to address urban traffic congestion and enhance the operational efficiency of signalized intersections. In a mixed traffic environment comprising both CVs and non-CVs, leveraging multi-resolution CV data for more precise prediction of future traffic conditions has become a critical factor in improving the effectiveness of traffic control systems. This thesis proposes a novel hybrid approach that seamlessly integrates traditional traffic flow models with deep learning techniques to predict and optimize traffic at signalized intersections with CVs. The proposed method first utilizes a Long Short-Term Memory (LSTM) Neural Network model based on CV data to predict the in-flow rates at intersections. Subsequently, a shockwave theory is applied to the predicted in-flow rates for accurate queue profile prediction. Finally, a signal optimization algorithm is developed to search for optimal phase sequences and durations within a forward time window to minimize vehicle delay. The simulation platforms are built for both a virtual and a real-world intersection, respectively, to evaluate the effectiveness of this hybrid approach in predicting queue profiles and optimizing signal timings. The results demonstrate that the proposed hybrid model performs well in predicting total delay at intersections under various market penetration rates (MPRs) of CVs and traffic demand levels. Furthermore, in a comparative analysis with the actuated control and fixed-time control, the proposed control algorithm is proven to outperform them significantly under medium and high demand levels. Under low demand levels, the proposed control algorithm has effectiveness similar to the actuated control but remains superior to the fixed-time control. Additionally, the results indicate that the effectiveness of this algorithm improves with higher MPRs and is still better than that of the actuated control, even under lower MPRs.