01659755

Component

Short-term travel time prediction generally utilizes real travel time values within a sliding time window to predict travel time one or several timestep(s) in future, which is significant for advanced traffic management and route planning. To maintain accuracy and reliability of prediction, it is necessary and challenging to capture the causal relationship of timesteps and prepare for the nonstationary properties and abrupt changes in travel time series. And a single prediction model is often inefficient in achieveing these requirements. Recent achievements of deep learning methods in regression and prediction shed a light on innovations of time series prediction. This study proposed an ensemble learning (or model combination) method based on convolutional neural networks (CNN) and long short-term memory neural networks (LSTM). In this method, 1) several base models are established and separately trained on train data to get their optimal prediction output on validation data; 2) a CNN-LSTM model is established for emsemble learning, which uses the prediction results of each base model on train data and validation data as input, and then the CNN-LSTM model is trained; 3) in the last stage the trained CNN-LSTM model is used for predicting test data. Finally, the proposed ensemble learning method is examined on a 90-day travel time dataset from Caltrans Performance Measurement System (PeMS), using 5 sets of predicting horizons. A series of benchmark models are also implemented for comparison, including linear regression, Ridge and Lasso regression, ARIMA and two kinds of deep learning models - deep neural network (DNN) models, and LSTM models. The results demonstrate the advantage of the proposed method on almost all scenarios, even outperformed individual deep learning model. This study can provide insights for optimizing deep learning model structures and for developing new ensemble learning method for traffic prediction.

This paper was sponsored by TRB committee ABJ30 Standing Committee on Urban Transportation Data and Information Systems.

18-06273

English

9p

2018

Transportation Research Board 97th Annual Meeting

Digital/other

Figures; References; Tables

Machine learning; Mathematical prediction; Optimization; Regression analysis; Time series analysis; Travel time

Highways; Planning and Forecasting

Transportation Research Board Annual Meeting 2018 Paper #18-06273

TRIS, TRB, ATRI

Jan 8 2018 11:37AM