Short-term travel time prediction by deep learning: A comparison of different LSTM-DNN models

Predicting short-term travel time with considerable accuracy and reliability is critically important for advanced traffic management and route planning in Intelligent Transportation Systems (ITS). Short-term travel time prediction uses real travel time values within a sliding time window to predict travel time one or several time step(s) in future. However, the nonstationary properties and abrupt changes of travel time series make challenges in obtaining accurate and reliable predictions. Recent achievements of deep learning approaches in classification and regression shed a light on innovations of time series prediction. This study establishes a series of long short-term memory neural networks with deep neural layers (LSTM-DNN) using 16 settings of hyperparameters and investigates their performance on a 90-day travel time dataset from Caltrans Performance Measurement System (PeMS). Then competitive LSTM-DNN models are tested along with linear models such as linear regression, Ridge and Lasso regression, ARIMA and DNN models under 10 sets of sliding windows and predicting horizons via the same dataset. The results demonstrate the advantage of LSTM-DNN models while showing different characteristics of these deep learning models with different settings of hyperparameters, providing insights for optimizing the structures.

I. Introduction

The short-term prediction of traffic flow parameters including flows, speeds, densities and travel time compose an important major part in development and application of Intelligent Transportation Systems (ITS), forming the foundation of Advanced Traffic Management Systems (ATMS) and providing essential supports for travel guidance and route planning of travelers. A time series of traffic flows or travel time generally has similarity among its inner sub-series corresponding to the same weekdays (e.g., Monday in last week to Monday in next week), weekends and same quarters. Despite the periodicity and seasonality, however, these time series are always being affected by various external factors including traffic crashes, fluctuation in traffic demand and abrupt changes of weather conditions, which results in nonstationary properties and brings difficulties to the prediction.