Spatial matrices for short-term traffic forecasting based on time series

Intelligent traffic systems require data and recent studies on Short-Term Traffic Forecasting (STTF) have incorporated the Spatiotemporal (ST) aspect to improve predictions. They consider road network characteristics by incorporating the impact of traffic at a given location and on its neighbouring locations. Spatial relationship weight matrices, or simply spatial matrices, facilitate direct incorporation by considering the correlation factors between various points. The Space-Time Autoregressive Integrated Moving Average Model (STARIMA) enables the application and comparison of various types of spatial matrices in ST-STTF. In this paper, we compare three types of neighbourhoods and eight types of weights to assess their impact in ST-STTF for different infrastructure and traffic characteristics. Several experiments in the road network of the city of São Paulo, Brazil, revealed that asymmetric matrices that consider the upstream and downstream traffic flow present higher accuracy than the commonly used symmetric matrices. The use of asymmetric matrices improved accuracy in 77.3% of scenarios, particularly when employing an asymmetric weight based on adjacent section speeds and travel time. Moreover, grouped matrices (GM) required less computational time to estimate the models when compared with contiguity (CN) and time lag (TL) matrices. Therefore, our results show the impacts of applying several spatial structures into short-term traffic prediction models and provide practical prediction methods for an urban network based on real traffic data, with a case study in one of the largest cities in Latin America.