Institute of Transportation, MOTC

       Travel time prediction is one of the main development projects to promote intelligent transportation systems in Taiwan, to solve air pollution and economic losses caused by traffic congestion. Researches in recent years often use a certain amount of data and different machine learning algorithms to make predictions. There are many factors that affect travel time, but previous studies only used single or a few features to predict. This research hopes to develop a real-time prediction methods, considering multiple features related to travel time in order to predict travel time between interchanges under current conditions, apply it on the segment of freeway which without signalize intersections. Two types of data, VD and ETC, are used in data collection, and estimated travel time information in a more accurate way. It takes into account that choosing k-NN method is about the prediction accuracy and the applicability of the data, and this research improve the method of distance calculation, estimating the predicted value and dynamic k. The results can be used as a reference for road authorities and users.
    This research is tested with the data of Freeway bureau, MOTC, the results show that the best feature vector combination we find can effectively improve the prediction accuracy compared with single feature vector, and it shows that the distance calculation method proposed in this research can also improve the prediction accuracy, in terms of the method of estimating the predicted value, the average method is better, but the dynamic k has limited effect on accuracy improvement. Under the best setting of this research, the average MAPE is 2.57%, MAE is 63.4, RMSE is 118.9, shows that it has the ability to accurately predict freeway travel time.