Predictive analytics for supply chain resilience in urban infrastructure networks using graph convolutional networks.

Abstract

Urban infrastructure networks underpin modern supply chain operations, yet their vulnerability to disruptions poses significant challenges to urban resilience; understanding and predicting such disruptions requires analytical frameworks that jointly leverage network topology and flow dynamics. This study proposes a three-stage framework integrating complex network analysis, Graph Convolutional Network (GCN)-based predictive modeling, and resilience stress testing, validated on a large-scale New York City dataset comprising 6,302 infrastructure nodes, 96,641 disruption events, and 100,000 traffic flow records. The proposed GCN model achieves an RMSE of 5.8 and R2 of 0.92, significantly outperforming ARIMA, SVR, LSTM, and XGBoost baselines, with ablation experiments confirming that graph structure reduces prediction error by 29.3%. Resilience analysis reveals that targeted hub-node attacks cause rapid initial fragmentation, while a crossover effect at approximately 30% node removal shows that random failures induce greater cumulative damage. Temporal factors and historical flow patterns are identified as the dominant predictors, collectively accounting for 65.5% of feature importance. The integrated framework bridges flow prediction and resilience assessment, enabling identification of dual-risk nodes for prioritized infrastructure protection and providing actionable insights for enhancing urban supply chain resilience.