Principal Component Analysis Embedded Temporal Attention Spatiotemporal Graph Convolutional Networks for Traffic Flow Prediction

Abstract

Precise and instantaneous traffic flow prediction is vital in intelligent transportation systems. To address the issues of data redundancy caused by sparse adjacency matrices in spatial modeling of traffic flow and insufficient capture of temporal dependencies in traffic flow, a spatiotemporal graph convolutional network (principal component analysis embedded spatiotemporal graph convolutional network [PESTGCN]) traffic flow prediction model with principal component analysis (PCA) embedded in temporal attention is proposed. The model consists of PCA, enhanced graph convolution (En-GCN), and temporal information extraction modules. PCA reduces the dimensionality of the data, eliminates redundant information in sparse matrices, and enhances the spatial representation capability of the matrix. En-GCN further captures the dynamic spatial correlations of traffic flows through adaptive variations of learnable parameters. The temporal information extraction is achieved through a stacked architecture of gated recurrent unit (GRU) layers, forming the main diagram-GRU (MD-GRU), aimed at capturing short-term temporal dependencies. Meanwhile, an attention gated temporal convolution module incorporating an attention mechanism is designed in parallel with the MD-GRU. It is mainly used to enhance the ability to capture long-term time dependency and improve the model performance. Experimental results on two publicly available datasets show that the forecasting accuracy of the PESTGCN outperforms the current popular state-of-the-art models in both cases and shows good prediction performance.