Forecasting post-disruption demand scenarios with transferable LSTM-GCN neural network to facilitate infrastructure recovery

Abstract

This proposed work studies the needs for accurate post-disruption demand forecasting to enhance the recovery of power and water systems. Unlike conventional methods that assume static demand patterns, we recognize that consumer behavior can change significantly after disruptive events. To capture these dynamic shifts, we develop a machine learning framework that integrates Long Short-Term Memory (LSTM) networks and Graph Convolutional Networks (GCNs). The LSTM component models the temporal evolution of demand based on time-series data, while the GCN component captures the spatial and structural interdependencies within the network. This combined approach enables adaptive, real-time forecasting of demands as the system conditions and consumer behaviors evolve during recovery. This study highlights the importance of dynamic demand estimation and the potential of machine learning models to enhance infrastructure resilience under extreme event scenarios.