Emergency relief materials are fundamental to effective earthquake response, and accurate prediction of their demand is critical for emergency rescue operations. Building on previous studies of factors influencing earthquake casualties, this paper selects eight key parameters—magnitude, occurrence time, damage intensity, seismic fortification intensity, affected population, population density, degree of building damage, and earthquake forecasting level—as predictive indicators. Gray correlation analysis is applied to examine the relationship between each factor and casualty numbers, and a hybrid model combining a radial basis function(RBF) neural network with a particle swarm optimization(PSO) algorithm is employed to predict casualties. Model comparison indicates that using all eight predictors achieves the highest accuracy. On this basis, a dynamic mechanism for forecasting emergency material demand is developed. Using the 2022 Luding earthquake as a case study, we forecasted initial, one-day, and three-day emergency supply requirements. The results demonstrate that the proposed dynamic forecasting workflow provides a robust methodological reference for scientific decision-making in the allocation of earthquake relief supplies and establishes a foundation for future research on demand prediction in disaster management.