Forecasting Train Arrival TIme Using Modular Artificial Neural Networks

The ability to accurately forecast train arrival times is essential for the safe and efficient operation of highway-railroad grade crossings (HRGC). Trains in the U.S. are required to give a minimum of twenty seconds of warning time before arriving at a HRGC. With the recent development of new detection equipment technology. detectors could potentially be employed further upstream of the HRGC which would result in earlier detection times. This information would be particularly useful for preemption strategies at signalized intersections located near the HRGC (IHRGC). For example, earlier warning times could be used to reduce or eliminate the risk of pedestrian movements at IHRGC being truncated in an unsafe manner. In this study, a modular Artificial Neural Network is used to forecast the train arrival time at a HRGC. An ANN was adopted because there is a non-linear relationship between the independent variables (i.e. train speed profile) and the dependent variable (i.e. arrival time at HRGC). A modular approach was used because the trains often have different characteristics depending on their cargo and the operational rules in effect at the time they are detected. The test bed was a railway corridor located in College Station, Texas. Doppler radar equipment was utilized to measure train speed and direction approximately 2.2 km from the HRGC. The arrival time forecast was performed 100 seconds after the train was first detected. Approximately 190 trains were used for training the ANN and 76 trains were used for testing. It was found that a modular architecture (with 4 categories) gave superior results to that of a simple ANN model, standard regression techniques, and simple forecasting methods. For example, the average absolute error was reduced by 29.6 percent as compared to the current method.