A Self-Imputing Deep Multitask Sequence Model for Traffic Disruption Detection in Extreme Conditions

Abnormal traffic patterns caused by extreme events can significantly disrupt traffic flow across large regions of urban road networks. Timely and reliable detection of such disruptions is crucial for effective traffic management. However, existing anomaly detection methods cannot simultaneously detect disruptions at both network and local levels, and they typically struggle with incomplete traffic data, which is a common issue during extreme events. To address these limitations, we propose a novel Self-Imputing Deep Multitask Sequence (SI-DMSeq) model for dual-level traffic monitoring and disruption detection. The model simultaneously performs two distinct tasks: (1) reconstructing sequences of traffic data, which enables anomaly detection through an autoencoder-based approach, and (2) making one-step-ahead predictions, which serves to impute missing data for the subsequent time step. To support robust training despite data gaps, we incorporate an iterative imputation strategy during the training phase. During inference, the reconstruction task is used to detect traffic disruptions at both network and local levels, while the prediction task provides reliable imputation of missing values under typical traffic conditions. We evaluate the SI-DMSeq model using historical traffic data from the Manhattan road network in New York City, focusing on disruptions caused by Hurricane Sandy in 2012 and the Christmas blizzard in 2010. The results demonstrate that the proposed method accurately identifies disruptions at both the network and local levels and efficiently imputes missing traffic data.