Traffic Profile Prediction, 10-R8548
David W. Vickers
Adam K. Van Horn
Matt D. Weatherston
Kenneth L. Holladay
Inclusive Dates: 04/01/15 – Current
Background — Advanced Traffic Management System (ATMS) software receives a continual stream of data from roadway sensors, police computer-aided dispatch systems, courtesy patrols, emergency vehicles, and its operators. As ATMSs have increased in complexity, they have generated and stored a large amount of traffic and event data. The daily operations of a traffic management center (TMC) focus on reacting to changing traffic patterns and incidents as they occur. There is a need to be able to predict more quickly traffic responses to planned events, quickly detect traffic anomalies, and alert the public faster and with more accuracy to the possibilities of travel delays. The goal of this project is to investigate the techniques necessary to change the TMC approach from the current "detect, respond, mitigate" to a proactive "predict, act, prevent." To do so, the ATMS must be able to predict traffic profiles from historical and current data. Predicted traffic profiles are made up of forecasts for the near future values (e.g. within the next 30 minutes) of traffic attributes such as speed, volume, and occupancy.
Approach — This project is experimenting with combining current data with historical data and inputs describing planned and unexpected circumstances (such as sports events, construction, weather, or accidents) to evaluate the ability to predict traffic profiles. Frequently careful analysis of the data is necessary to correct for abnormalities and artifacts introduced into the data. Figure 1 illustrates a sudden change in the values of the volume data when a configuration was changed. The data after the sudden change in the values must all be divided by 10 to be correct. Figure 2 illustrates an interactive visualization of traffic data (in this case vehicle speed during a 24-hour period along a segment of I-4) along with events (the vertical bars) that impact the data. The objective of the research is to evaluate the feasibility of adding traffic profile prediction capabilities to SwRI supported ATMS software. Among the goals of the project are:
- Testing baseline traffic profile prediction algorithms using recent data.
- Evaluating algorithms for adding offsets to the recent data baselines using periodic functions of traffic.
- Testing mechanisms for detecting abnormal conditions.
- Characterizing the impact of the attributes of detected conditions.
- Evaluating the effectiveness of algorithms for integrating the impact of abnormal conditions into traffic profile predictions.
- Measuring the improvements in traffic profile predictions from combining current data, periodic historical data, and abnormal traffic condition data as a basis for prediction.
- Validating the utility of the combined algorithm.
Accomplishments — Raw data, rolled up data, and other traffic information have been retrieved from files, databases and websites so that they can be combined and used for traffic profile prediction. Figure 3 shows the segment of I-4 from which the initial data was analyzed. The stars represent the locations of the detectors. An autoregressive integrated moving-average (ARIMA) model has been developed as a baseline for traffic profile prediction. Figure 4 illustrates the predictions of the ARIMA model selected. Data for I-4 in Orlando has been analyzed across lanes, in both directions and across various time periods (days, weeks, months, years). Sensor data has been analyzed and, where necessary, corrected to address anomalies found. Visualizations of traffic parameters, such as speed, volume, and occupancy across a set of detectors and time have been used to evaluate the impact of periodic and abnormal events. Data for different days of the week and different seasons, as well as those influenced by events, have characteristic shapes. Figure 5 shows the typical shapes (called medoids) of traffic speeds in a specific lane at a specific location. This type of analysis is called partitioning around medoids (PAM) and is useful in finding typical shapes from a collection of shapes. Analyses of the impact of rolling the data up to various time spans and impact of the length of data considered in making the models have begun.