Analyzing Taiwanese traffic patterns on consecutive holidays through forecast reconciliation and prediction-based anomaly detection techniques

TL;DR

This paper presents a simple, efficient prediction-based method for detecting traffic anomalies on Taiwanese highways during holidays, capturing seasonality and spatial autocorrelation to inform traffic management.

Contribution

It introduces a reconciliation-based OLS forecasting approach with bootstrap prediction intervals for anomaly detection in complex traffic datasets.

Findings

Detected high anomaly rates on specific highways and regions during holidays.

Identified spatial and directional variations in traffic anomalies.

Demonstrated the method's effectiveness in modeling complex traffic patterns.

Abstract

This study explores traffic patterns on Taiwanese highways during consecutive holidays and focuses on understanding Taiwanese highway traffic behavior. We propose a prediction-based detection method for finding highway traffic anomalies using reconciled ordinary least squares (OLS) forecasts and bootstrap prediction intervals. Two fundamental features of traffic flow time series -- namely, seasonality and spatial autocorrelation -- are captured by adding Fourier terms in OLS models, spatial aggregation (as a hierarchical structure mimicking the geographical division in regions, cities, and stations), and a reconciliation step. Our approach, although simple, is able to model complex traffic datasets with reasonable accuracy. Being based on OLS, it is efficient and permits avoiding the computational burden of more complex methods. Analyses of Taiwan's consecutive holidays in 2019, 2020, and 2021 (73 days) showed strong variations in anomalies across different directions and highways. Specifically, we detected some areas and highways comprising a high number of traffic anomalies (north direction-central and southern regions-highways No. 1 and 3, south direction-southern region-highway No.3), and others with generally normal traffic (east and west direction). These results could provide important decision-support information to traffic authorities.