Predicting non-recurrent congestion impact: A pattern-based approach for speed drop ratio prediction using weighted K-nearest neighbors

Abstract

Traffic congestion remains a major challenge in developed countries, disrupting mobility and affecting economic and social activities. Among its various types, non-recurrent congestion - caused by unexpected events such as accidents, maintenance, or debris - remains difficult to predict due to its irregular spatiotemporal dynamics. While existing models effectively forecast recurrent traffic, they are less applicable to non-recurrent events characterized by abrupt and anomalous patterns. This study presents a pattern-based framework that integrates the weighted K-nearest neighbor (WK-NN) algorithm with dynamic time warping (DTW) for similarity-based prediction of non-recurrent congestion impact. The framework estimates speed drop ratios (SDRs) and propagates the predicted effects to neighboring road segments, enabling a network-level assessment of disruption. By identifying historical patterns most similar to the current incident, the proposed approach enhances interpretability and traceability for operational use. We evaluate the method using 2780 real-world traffic incident records combining data from the Korean National Police Agency and NAVER Corporation. Experimental results demonstrate that the proposed framework achieves consistent and competitive performance compared with benchmark machine learning and deep learning models. These findings suggest the framework's potential for supporting practical decision-making in traffic control centers through timely and interpretable congestion impact forecasts.