Temporal Aggregation Effects in Spatiotemporal Traffic Modelling.

Area of Science:

• Urban planning and transportation science
• Data science and time series analysis
• Geospatial modeling and traffic engineering

Background:

• Spatiotemporal models are widely used for urban traffic forecasting.
• Model specification, particularly temporal aggregation of traffic sensor data, significantly impacts spatial structure and forecasting accuracy.
• Current practices often involve arbitrary selection of temporal aggregation levels, potentially leading to suboptimal model choices.

Purpose of the Study:

• To investigate the impact of temporal aggregation levels on the forecasting performance of various spatiotemporal models.
• To compare different spatiotemporal model specifications, including vector autoregressive models and classical time series models.
• To provide empirical evidence for selecting optimal temporal aggregation levels for urban traffic forecasting.

Main Methods:

• Extensive experiments using real-world traffic sensor data.
• Analysis across multiple dimensions: temporal aggregation, forecasting horizons (one-step, multi-step), spatial complexity, spatial restriction methods (unrestricted, travel-time-based, correlation-based), and series transformation (original, detrended).
• Comparison of travel-time-based and correlation-based spatially restricted vector autoregressive models against univariate and multivariate time series models.

Main Results:

• The temporal aggregation level plays a critical role in identifying spatiotemporal traffic flow structures and selecting appropriate model specifications.
• Arbitrary selection of temporal aggregation levels can result in incorrect conclusions about optimal model performance.
• Empirical results highlight the influence of temporal aggregation on forecasting accuracy for different spatiotemporal models and horizons.

Conclusions:

• The choice of temporal aggregation level is a crucial, non-arbitrary decision in spatiotemporal traffic forecasting model specification.
• Future research should validate model specifications across various temporal aggregation levels to ensure robust forecasting.
• Methodologies for urban traffic forecasting should incorporate the systematic evaluation of temporal aggregation levels for optimal model selection.