Short Term Traffic Flow Prediction with Spatial-Temporal Neural Networks on England Highway

Published:

This project is initially submitted in partial fulfillment of the requirements for the MRes degree in Advanced Computing of Imperial College London. Accepted as a poster in MLImperial2017.

Abstract

Both temporal and spatial features provide significant implications for short-term traffic volume prediction. The problem is challenging due to various non-linear temporal dynamics at different locations, complicated spatial dependencies and difficulty for longer-step ahead forecasting. We propose two deep learning models, CNNLSTM with attention mechanism (CNN-LSTM-Attn) and Temporal-Spatial-LSTM (TSLSTM) to incorporate temporal and spatial correlations. Experiments show that both models outperform baselines on the Highways England dataset and the CNN-LSTMAttn achieves lowest MAPE 9.26% on 2-hour traffic volume prediction. We also evaluate the CNN-LSTM-Attn on the KDDCUP17 dataset and our model defeats the model that got first place in the competition with lower MAPE 10.48%. Our models achieve 2-hour forecasting, which is longer than previous literature, with outstanding accuracy and robustness.

Demo

England Highway Prediction Visualisation (GIF)


Visulisation on Global Data Observatory

Files

Poster Thesis