Multi-Level Disruption Detection in Road Networks Using LSTM and MCUSUM: A Case Study of Hurricane Sandy

Real-time monitoring of road network traffic patterns and detecting disruptions is crucial for intelligent transportation systems. This study proposes a multi-level framework to detect traffic disruptions at both the network-wide and road segment levels. The method uses long short-term memory (LSTM) neural networks to predict hourly traffic speeds, with prediction errors serving as disruption indicators. Network-wide anomalies are detected using a multivariate cumulative sum (MCUSUM) control chart, while road segment-level disruptions are identified by decomposing the MCUSUM statistic using the correlation-maximization (Corr-Max) transformation. Applied to Manhattan’s road network during Hurricane Sandy in 2012, the method demonstrated high accuracy in detecting both macroscopic and microscopic traffic anomalies. Unlike traditional approaches focusing solely on network-wide or segment-specific disruptions, this framework provides comprehensive insights, enabling adaptive traffic management systems to enhance network resilience and emergency response capabilities.