In response to this request, NCAR applied machine learning algorithms to predict travel times on road segments for two minute time intervals. The machine learning algorithms were applied to a database consisting of weather data including temperature and precipitation observations from nearby MADIS stations in conjunction with average travel time observations in order to create route-specific forecasting models. The forecasting models were then used to produce 15-minute travel time forecasts over the next hour, and then hourly forecasts out to 24 hours. The road segment travel times were then combined to create travel times for the individual routes. Here the individual routes range from 1 mile up to 116 miles in length and consist of one or more individual road segments.
This presentation will outline a number of the many challenges that were faced in developing and running the forecasting system, from quality controlling historical data to handling changes in traffic patterns. It will discuss the methodology used to address these challenges as well as the algorithm for combining the segment specific travel time forecasts into a total route forecast. In concluding the presentation, performance results and ideas for future development will be discussed.