Neigbourhood Embedding for Improving Diffusion based Short Term Traffic Flow Prediction

Abstract

This paper describes a short-term traffic flow forecasting approach that combines efficient probability based diffusion mechanism and embedding of surrounding information to handle both ordinary situation and abnormal situations. A discrete diffusion model is insufficient in handle Spatial temporal traffic flow data sampled at regular time interval as there may be information loss during volatile environment. The hybrid model utilize DCRNN as discrete diffusion model and GRU as embedding, combining both for accurate prediction. Preliminary results shows improvement in both microscopic and macroscopic scales indicating the potential of hybrid approach towards accurate and efficient short term traffic flow forecasting