A parallel-res GRU architecture and its application to road network traffic flow forecasting

Gated Recurrent Units (GRU) is an effective architecture on time series analysis. But deeper GRU models are more difficult to train and degrade rapidly. In this paper, we proposed a Parallel-Res GRU architecture for road network traffic flow forecasting. We explicitly reformulate the Parallel-Res Path as learning residual functions with reference to the input layer and output layer, instead of stacked plain GRU layers. And we construct multi-level residual architecture. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can improve accuracy from considerably increased depth and reduce the degradation. On UCI's PEM-SF dataset, we construct a 20 layers Parallel-Res GRU model for road network traffic flow forecasting, and the experiments indicate a considerable improvement comparing with normal stacked GRU models.