Research on material demand prediction based on ensemble learning algorithm

Natural disasters such as earthquakes pose serious threats to the human society. As crucial material foundations for emergency response, the degree to which the supply of emergency relief materials matches the demand directly determines the rescue effectiveness. The uncertainty caused by multiple factors makes accurate predictions of demand for emergency living materials an urgent requirement. To address the practical needs of demand prediction and the limitations of existing research, this study focuses on post-earthquake living material demands. By comprehensively considering historical data, disaster types, and affected areas, and leveraging the advantages of multiple machine-learning models, we developed a model called GS-RFGBXGB-LR based on a stacking ensemble strategy. Using case data of earthquakes above a magnitude of 5 in Yunnan Province, China, since 2007, comparative analyses with baseline models (K-nearest neighbors, support vector machine, and back-propagation neural network) have demonstrated that the proposed model achieves a 29.65%–89.06% improvement in the mean absolute percentage error, showing higher prediction accuracy. This method provides valuable support for emergency decision-making.