Anomaly Detection of Road Ranking Shifts Due to Traffic Accidents Using Deep Learning on Time Series Data

Abstract

This study developed an anomaly detection model based on Long Short-Term Memory (LSTM) autoencoders to identify abnormal shifts in road ranking scores caused by traffic accidents in Magelang, Indonesia. Road rankings were derived from time-series data of traffic indicators collected between 2015 and 2020, including volume-to-capacity ratios, heavy vehicle proportions, and average speed. The model was trained on non-accident data to learn normal traffic behavior and subsequently detect deviations. Anomalies were identified when reconstruction errors exceeded statistically defined thresholds and were evaluated against verified accident records. The model achieved a precision of 82%, recall of 75%, and an AUC-ROC of 0.87, demonstrating strong performance in detecting significant disruptions, particularly severe accidents involving fatalities or serious injuries. Analysis showed that detected anomalies were concentrated on high-risk roads and during peak traffic hours. These findings highlight the potential of LSTM-based models for integration into intelligent transportation systems to support real-time accident detection and proactive traffic management in developing urban environments.