Texas Study Showed Machine Learning-Based Travel Time Predictions Were 40 Percent More Accurate During Peak Periods Compared to Traditional Approaches.
Field Testing Conducted Along Corridors in Austin and El Paso Evaluated Machine Learning-Based Travel Time Prediction Models.
El Paso
Austin, Texas, United States
An Exploration of the Use of Artificial Intelligence for Enhanced Traffic Management, Operations and Safety
Summary Information
Artificial intelligence (AI) has a growing role in improving the efficiency, safety, and coordination of modern transportation systems. The Texas Department of Transportation (TxDOT) conducted a project focused on applying AI and machine learning (ML) techniques to support Integrated Corridor Management (ICM), a promising strategy designed to mitigate congestion and improve operational coordination in urban transportation networks. Key activities included a literature review, stakeholder workshop, survey, development of prototype ML models for four high-priority use cases, and field testing of short -term travel time prediction models. Prototype models were built using various data sources and ML techniques to evaluate feasibility across applications, including safety analysis, traffic patterns, real-time signal control, and travel time prediction. A framework was also developed to support model training, testing, and potential deployment within TxDOT’s network.
METHODOLOGY
The prototype model development involved integrating emerging and traditional data sources, such as connected vehicle event data, safety data, probe vehicle speeds, and TxDOT’s ITS traffic volumes within supervised, unsupervised, and reinforcement learning models. The project also utilized microsimulation to explore real-time signal control strategies. Field testing of short -term travel time prediction was conducted over multiple corridors in Austin and El Paso, using one to two years of historical data from 2019 to 2022. The evaluation scope included both segment-level and route-level predictions, using time series and ML models. This use case included historical speed and travel time data, in addition to traffic volume and occupancy data. The study focused on the effectiveness of the prediction models compared to traditional approaches as the evaluation metric.
FINDINGS
- Short-term travel time predictions using the ML-based models were up to 40 percent more accurate than those obtained from a naïve method that simply added travel time along all segments at the time that the trip started.
- ML-based models showed a decrease in error rate to 10 percent from 15 percent in providing comparative travel times across routes.
