Enhanced Self-node Weights Based Graph Convolutional Networks for Passenger Flow Prediction

Abstract

Accurate and real-time passenger flow prediction is of great significance for realizing intelligent transportation systems. However, due to the complexity and unstable change of traffic network passenger flow data, passenger flow prediction remains a challenging problem in transportation research field. Moreover, the core problem is how to obtain the spatial and temporal characteristics efficiently. In this paper, we propose an Enhanced Self-node Weights Based Spatial-Temporal Graph Convolutional Networks (EST-GCN) model to capture the spatial and temporal characteristics. Specifically, in order to capture the spatial characteristics, we optimize the ability of Graph Convolutional of Network (GCN) in extracting the spatial characteristics of rail transit networks based on the difference maximization of aggregated information, hoping to solve the problem that GCN cannot fit peak value accurately. As for temporal characteristics, we leverage the Gate Recurrent Unit (GRU) model to obtain dynamic changes of passenger flow data to capture them. The EST-GCN model is a combination of these two models. Based on the Shanghai dataset, we use the proposed EST-GCN model for simulation experiments, and compare our proposed method with other mainstream passenger flow prediction algorithms. The experimental results demonstrate the superiority of our algorithm.