Crash Prediction Model for Tennessee Showed Response Time Savings, with a 19 Percent Average Improvement When Roadside Assistance Trucks Were Available.
Simulation Study Illustrated How the Tennessee Highway Patrol Could Apply a Predictive Model to Aid in Resource Planning Statewide.
Tennessee, United States
Predictive Analytics for Traffic Management Systems
Summary Information
Predictive analytics develops and applies mathematical models to make statements about the future state of a system. The Tennessee Department of Transportation (TDOT), in collaboration with the Tennessee Department of Safety & Homeland Security and Vanderbilt University, developed a predictive analytics tool called Crash Reduction Analyzing Statistical History (CRASH). By predicting the likelihood of crashes, the tool could be used for resource allocation (staff per shift, highway safety patrol routes) and emergency response planning (extreme weather and special events). Patrols can be assigned where and when the model predicts the risks of serious crashes are highest.
METHODOLOGY
The CRASH tool used commercially available statistical software, historical crash data from the Tennessee Integrated Traffic Analysis Network, National Oceanic and Atmospheric Administration weather data, and special events data (e.g., sporting events, holidays, and festivals). Specifically:
- Incidents reported by TDOT, February 2017 to May 2020.
- Weather data from 40 weather stations across the state, February 2017 to June 2020.
- GPS probe traffic data (congestion, free flow speeds, and confidence scores) from April 2017 to December 2020.
The predictive model generated crash forecasts by 4-hour temporal, day, week, and 42 square mile geographic blocks.
FINDINGS
- The model showed a 19 percent average response time improvement when 20 roadside assistance trucks were available.
- There was up to an 8 percent average improvement in response times for multiple different numbers of trucks.
- When the number of trucks was capped at 10, the model decreased the average travel distance by responder per accident by up to 4.5 km (about 2.8 miles), resulting in a travel time savings of more than 4.5 minutes.
- The model reduced the total number of unattended accidents to up to 75 percent and 50 percent for the mean and maximum number of unattended accidents during a 4-hour time window.
- The model was 70 percent accurate in identifying areas of concern for alcohol, drug, and crash-involved incidents.
Additional Source(s):
- "Collaborative Research Project to Coordinate the Data from the CRASH Predictive Analytics Program Between TDOT and TDOSHS" (2021). Source URL: https://rosap.ntl.bts.gov/view/dot/61069 (accessed 07/30/2025).
