Extending Hybrid GRG-NS With LSTM-Based Demand Forecasting for Dynamic Multi-Depot Routing in Disaster Logistics

Abstract

Disaster logistics management requires accurate demand forecasting and efficient routing optimization to ensure timely distribution of emergency supplies under dynamic and uncertain conditions. Conventional routing approaches often experience limitations in handling fluctuating disaster demand, resulting in inefficient distribution performance and increased operational costs. This study proposes an integrated LSTM–Hybrid Generalized Reduced Gradient and Neighborhood Search (LSTM–Hybrid GRG–NS) framework for disaster-demand forecasting and routing optimization. The proposed approach combines Long Short-Term Memory (LSTM) for sequential demand prediction with a hybrid GRG–NS optimization mechanism to improve routing efficiency and solution convergence. Experimental evaluation was conducted using disaster-demand scenarios and routing datasets to assess forecasting and optimization performance. The forecasting results demonstrated strong predictive capability with low MAE, RMSE, and MAPE values, indicating that the LSTM model effectively captured temporal demand patterns. Furthermore, the routing optimization results showed that the proposed framework successfully generated stable and near-optimal routing solutions while maintaining full demand fulfillment and efficient vehicle utilization. The convergence analysis also confirmed that the optimization process converged consistently within a limited number of iterations. Overall, the proposed LSTM–Hybrid GRG–NS framework provides an effective and reliable decision-support approach for proactive humanitarian logistics and disaster-routing management.