Deep learning approaches for traffic prediction

Abstract

The rapid and continuous population growth and urbanization movements have resulted in the increase of number of vehicles on the road. Some detrimental effects on people's life are observed such as lost of productivity, air pollution and higher fuel consumption. There is a crucial need for intelligent transportation systems (ITS).

A well designed ITS can support many functions such as route planning and traffic management. They both rely on an accurate traffic prediction. In this thesis, we focus on the data-driven models for traffic prediction task. There are many existing works in this field, however they either require strong assumptions or are shallow in the model structure. Therefore, we explore the problem with deep learning approaches which demonstrate remarkable capability in extracting deeper representation in the data. Since traffic prediction depends on the historical values and is essentially a time series forecasting problem, we firstly investigate the temporal correlations among the data with LSTM networks on a one-step traffic flow prediction task.

Next we extend the task to multi-step traffic speed prediction task. However there are some potential issues in naive LSTM approach. It has difficulties for longer step decoding by relying on a compressed fixed-length vector solely. Also, the recurrent connection makes the computation very costly. Therefore we proposed a full attention model which discards the recurrent connection and uses attention to assist the fixed-length vector during decoding.

On the other hand, road conditions are always affected by their neighboring roads. Therefore we investigate the spatial dependencies among the neighboring nodes and proposed a graph diffusion recurrent neural network to incorporate both temporal and spatial features.

All three proposed models are verified in real world traffic dataset. The LSTM model is tested on the PeMS dataset which we crawled the data from a website of US government project. We conducted experiments for the other two models on METR-LA dataset, which becomes popular recently and is used as a standard dataset for traffic prediction task. All approaches have demonstrated better performance in terms of MAE, RMSE and MAPE than strong baseline models.