Vehicle speed forecasting based on GCN-LSTM combined model

If exists any questions, please contact me via email the phone number!).

Format, structure and content requests are in the readme.docx file.

The dynamics of many real-world phenomena are spatio-temporal in nature. Traffic forecasting is a quintessential example of spatio-temporal problems for which we present here a deep learning framework that models speed prediction using spatio-temporal data. The task is challenging due to two main inter-linked factors: (1) the complex spatial dependency on road networks, and (2) non-linear temporal dynamics with changing road conditions.

To address these challenges, here we  explore a neural network architecture that learns from both the spatial road network data and time-series of historical speed changes to forecast speeds on road segments at a future time. In the following we demo how to forecast speeds on road segments through a `graph convolution` and `LSTM` hybrid model.  The spatial dependency of the road networks are learnt through multiple graph convolution layers  stacked over multiple LSTM,  sequence to sequence model, layers that leverage the historical speeds on top of the network structure to predict speeds in the future for each entity.

Posted in Uncategorized