E-STGCN: Enhanced Spatial-Temporal Graph Convolutional Network for Road Traffic Forecasting

Abstract

The paper proposes an enhanced spatial-temporal graph convolutional network (E-STGCN) for traffic flows forecasting. A spatial-temporal graph convolutional network was improved by introducing an additional criterion assessment that takes into account the mutual influence of traffic flows on adjacent sections of the road network. Additional criterion assessment proposes enhancements on the adjacency matrix construction, so the emphasis is placed on a section of the road network with its real characteristics and connections with other sections, which takes into account the direction of movement, ramps, turns, and geometric characteristics of the sections. Besides this, the adjacency matrix of E-STGCN reflects not only topological relations but also functional, logical, etc. Also, spatial-temporal components of the E-STGCN are based on the hourly, daily, and monthly elements of the traffic flow rather than the weekly element of a spatial-temporal graph convolutional network. The effectiveness of the proposed E-STGCN was evaluated in comparison with the LSTM network and STGCN on the CityPulse dataset. The proposed E-STGCN showed a lower prediction error than the LSTM network and STGCN.