Research Lead: Michael Zhang
UC Campus(es): UC Davis
Problem Statement: High occupancy vehicle (HOV) facilities allow carpool vehicles to bypass congestion and save travel time. However, they are often underutilized or overutilized, leading to either the waste of precious road capacity or the loss of travel time advantage over general purpose (GP) lanes. To address this issue, high-occupancy toll (HOT) lanes have gained popularity in recent years. HOT lanes allow single-occupant vehicles (SOVs) to pay to use the HOV lanes to make use of the spare capacity. When properly priced, HOT lanes preserve the advantage of HOV lanes, which encourages carpooling and thereby reduces congestion and vehicle miles. Currently, more than 100 miles of HOT lanes are operating nationwide. California has the largest HOT lane system, including the I-15 FasTrak section in San Diego and I-580/I-680 sections in Alameda County. Compared with HOV lanes, HOT lanes better utilize road capacity and offer SOVs more choices. Real-time pricing plays a key role in balancing the travel needs of carpoolers and SOVs and maintaining the travel advantage of HOT lanes. This requires pricing schemes that 1) take into account the price sensitivity of SOVs to travel time, 2) predicts the spare capacity of the HOT lanes, and 3) considers the operating conditions of GP lanes. Current pricing schemes are either simple schemes typically based on the traffic volume or average speed on the HOT lanes, without considering traffic conditions on GP lanes, or black-box schemes known only to private vendors operating the HOT lanes for profit.
Project Description: This project will adopt a data-driven approach to develop a new HOT lane pricing scheme that considers all three elements by mining existing data to reveal the underlying patterns between the traffic flow (on both HOT and GP lanes) and the toll rates, which are termed “flow-tolling patterns.” Traffic flow on HOT and GP lanes and their interactions are highly nonlinear and complex processes that are challenging to model using traditional modeling approaches. Data-driven approaches, such as machine-learning (ML)-based methods, on the other hand, offer a promising alternative to capture the underlying patterns and variations of such complex processes when provided with adequate amounts of data, and they are mostly assumption free. The I-580 Express Lane tolling dataset provided by the Alameda County Transportation Commission (CTC), which contains detailed information on posted toll rates, trip transactions, and traffic data for both HOT lanes and GP lanes, provides good opportunities to develop a deep sequence-learning model to extract underlying flow-tolling patterns. A control framework will be developed based on it to derive optimal pricing in real-time, with the aim of encouraging more SOVs to choose HOT lanes with reasonable charges while preventing HOT lanes from being congested. The optimal price will be input into the data-driven model to simulate how traffic responds to the price. The benefits of such a bi-level framework combining data-driven methods and optimal control are threefold: 1) it makes use of data-driven models to analyze the non-linear relationship between the current tolling scheme and traffic dynamics; 2) the data mining process is not restricted by the number of HOT lanes; and 3) based on the mined flow-tolling patterns, our real-time tolling control strategy will improve the current pricing mechanism. The proposed methodology is divided into two parts: 1) capturing the underlying variation of the traffic flow (on both HOT and GP lanes) with the dynamic toll rates; and 2) improving the existing tolling mechanism through controlling approaches, with multiple objectives that include improving HOT lane utilization and social welfare.
Status: In Progress