Performance evaluation of short-term time-series traffic prediction model

One of the key functions of an effective Advanced Traveler and Management Information System is the ability to model short-term predictions of traffic conditions on major freeways and arterials with reasonable accuracy. The impetus of forecasting traffic information rather than relying on real-time information is to allow the traveling public to be proactive in their travel decisions at both pretrip planning stage and en route. Traffic information is usually reflected by the traditional performance measures such as travel times and delays. Such measures are well perceived by the traveling public and, hence, are adopted in most short-term traffic prediction studies. When disseminated to the traveling public, forecasted travel time information will help the travelers make better trip decisions in terms of departure time, route selection, and mode selection, when transit becomes competitive. Several traffic prediction models have been developed in the past using a wide spectrum of modeling techniques. Most studies showed that prediction accuracy is often compromised by the underlying mechanism of prediction methods more than other influencing factors. The main objective of this study is to investigate the factors that have a significant impact on the forecasting accuracy of travel times using a nonlinear time series traffic prediction model. The ultimate goal is to identify the operational settings and the anticipated forecasting accuracy of the model before it can be fully implemented. The study was conducted using extensive amount of real-time data collected from the 62.5 km corridor of Interstate-4 in Orlando, Florida. Various scenarios were generated from a combination of model parameter settings and different traffic conditions in order to test the performance of the model. Relative travel time prediction errors were used as a measure of performance and were based on speed predictions at freeway detector stations. Statistical analysis was conducted to identify the parameters with a statistically significant effect on the model's performance.