A Passenger Flow Prediction Model Based on Graph Convolutional Network with Multivariate Spatio-temporal Correlation

Abstract

 Accurate prediction of short-term passenger flow is very important for rational planning and stable operation of cities, however, the problem of passenger flow prediction faces many challenges, including both the establishment of an effective spatio-temporal dynamic model structure and the necessity to comprehensively consider a variety of factors affecting the explicit and implicit passenger flow. So, a Multi-Variate Spatio-Temporal Correlation Graph Convolutional Network model (MVSTCGCN) is proposed. The model utilizes three kinds of spatially correlated graphs to construct a base graph, which is combined to capture spatio-temporal features globally; temporal attention mechanism, spatial attention mechanism, graph convolution operation, and spatio-temporal convolution constitute the spatio-temporal graph convolution module to capture local spatio-temporal features; meanwhile, the core module of graph convolution network is improved by being integrated wavelet transformation operators. The model is validated by New York taxi YellowTrip dataset and self-built dataset respectively; the simulation experiments show that the performance of our algorithm has more obvious advantages compared with other excellent algorithms.