Deep Learning Techniques for Traffic Speed Forecasting with Side Information

Abstract

Traffic speed prediction is an ongoing challenge for researchers, transportation agencies, and navigation applications. Involving real-world speed data makes the prediction complex and dynamic. The stochastic nature of traffic makes predictions using traditional statistical methods unsatisfying in terms of accuracy and performance. Recently, deep learning methods have gained more attention to capture this chaotic characteristic. This study conducts an encoder-decoder sequence to sequence learning manner and WaveNet with a side information model and compares the results with Autoregressive Integrated Moving Average. Using Waze crowdsourced speed data collected from 31 segments of Interstate 40 (I-40) in Tennessee, the proposed algorithms are trained and tested for short- and long term speed prediction (time steps from 5-minutes to 2-hours). Our experimental results demonstrate the WaveNet model with side information achieves the best performance with MAPE 4.40% for 5-minuets and MAPE 5.58% for 2-hours prediction.