MuSeFFF: Multi-stage feature fusion framework for traffic prediction

Abstract

A typical application of spatio-temporal data is traffic flow prediction. Precise traffic prediction needs to exploit the latent spatial, temporal and spatio-temporal dependencies. Most of the recent works on traffic prediction, based on deep learning models such as encoder-decoder, CNN, RNN and graph-based, fail to harness the spatial and temporal dynamics embedded in the data independently and accommodate the complex inter-dependency between location and time instances and their variability. This work proposes a novel deep learning Multi-Stage Feature Fusion Framework (MuSeFFF) for a futuristic traffic prediction in a road network. The latent correlations within the spatial and temporal features are extracted individually. At each stage of the framework, the extracted features are fused with the output of the ST-Conv unit and given as input to the successive ST-Conv units to address the stated problem. MuSeFFF has been trained and evaluated with the PeMS-BAY traffic dataset and compared with existing models to assess its performance. The proposed model outperforms other state-of-the models with 1.25% and 2.76% of improvement for medium and long-term prediction, respectively.