SGRU: A High-Performance Structured Gated Recurrent Unit for Traffic Flow Prediction

By incorporating attention The scenario of multivariate time series occurs in various mechanisms into GCN, it is possible to distinguish the domains of life. Researchers utilize historical weather data importance of different nodes and utilize Gated Recurrent from different regions to predict future rainfall intensity [1]. Units (GRU) for frequency domain feature extraction on long Taxi data is used to help cities predict travel resources time series, leading to a substantial improvement in prediction and reduce traffic congestion [2]. Exception monitoring is accuracy [7]. The aforementioned work demonstrates the performed on various states in manufacturing systems and effectiveness of GCN in learning spatial features and its strong internet services [3]. Since the introduction of Graph Convolutional adaptability to non-Euclidean structured data. Networks (GCN) in 2017 [4], this method has been widely used in the field of Multivariate Time Series (MTS) A. Problem 1: Dilated convolutions break adjacent time steps for spatial semi-supervised and self-supervised learning. By Focusing on the direction of traffic flow prediction (Figure interleaving one-dimensional convolution with gated linear 1,2), recent Temporal Convolutional Networks [8] and units (GLU) and graph convolution, and appending an output Graph WaveNet (GWNet) [9] have adopted the mechanism layer after this "sandwich" structure [5], accurate prediction of GCN. However, in the frequency domain, they use dilated of traffic flow speed can be achieved.