| Summary: | 碩士 === 國立交通大學 === 資訊科學與工程研究所 === 107 === In the development and application of intelligent transportation systems, accurate prediction of travel time is an indispensable application, so that the public can make better trip planning before travel begins. This helps to reduce passenger anxiety and transportation costs. Travel time estimation and prediction is a complex and challenging task, which tends not to be able to grantee the accuracy of prediction due to different vehicle combinations and external factors such as weather and events. The purpose of this study is to forecast the short-term national highway traffic on expressways, using the available national highway traffic information. In addition to travel time, our forecast target including the time needed for a congestion to resolve. For prediction features, we used date/time related information, primitive VD data (speed, flow, and road occupancy), and new features generated from VD data. Using the Gradient Boosting Regression Tree (GBRT) to predict, we propose a method to adjust the GBRT parameters and perform feature selection to improve the prediction accuracy. Observing the Pearson correlation coefficient to include optimal adjacent VD set in our prediction model, and using the quantile loss function can reduce travel time prediction error by 2.44% in average and the average MAE of TTF prediction by 0.21. Moreover, we adjust the GBRT hyper-parameters to improve the prediction accuracy. Our experiments show that the impacts of upstream and downstream traffic are not the same for different road segments. Therefore, we need to find the best adjacent VD combination for each road segment. The total congestion road length generated by our feature engineering helps to reduce the TTF prediction errors in three out of four selected routes. Feature transform is also performed on the selected road segments, so that the MAPE of TT prediction is reduced by 0.5% on one selected. Finally, comparing with the deep learning prediction errors, our GBRT model performs better on heavy traffic segments; it has a lower prediction error (the reduction is up to 1.08% in MAPE) and a smaller number of prediction errors at the time when the congestion is about to resolve.
|
|---|
