Predicting emergency supply demand during flood disasters is extremely challenging: available observation data are scarce, and the disaster’s evolution is highly dynamic, nonlinear, and uncertain. Meanwhile, existing studies have not considered the attrition rates of certain types of supplies, leading to a situation where some supplies are in excess while others are in shortage. To address these problems, this study proposes an adaptive ensemble learning framework for the statistical forecasting of emergency supply demand in few-shot scenarios. First, an adaptive dynamic weighting strategy is developed to integrate three complementary forecasting models, namely an improved GM(1,1) model and an enhanced grey incremental model for capturing short-term dynamic variations, together with the relatively stable improved Holt linear trend model. Second, the integrated forecasting model is employed to estimate the number of affected people during the disaster’s evolution. Third, by incorporating supply categories and attrition rates, a dynamic forecasting model for emergency supply demand is constructed to support a more reliable balance between supply and demand throughout the response process. The proposed framework is validated through case data. Experimental results show that the method achieves strong adaptability and predictive accuracy under limited data conditions. (1) In terms of predicting the number of affected people, compared with benchmark models, the proposed method reduces the mean relative error (MRE) by 2.19% and the root mean square error (RMSE) by 2.2726, while the R2 value is closer to 1. (2) Analysis of demand forecasts shows that the demand for Category I first increases slowly and then surges rapidly in response to the disaster’s evolution, consistent with the characteristics of disaster dynamics. The demand for Category II, due to every-other-day distribution and low attrition rates, exhibits a pattern of an initial surge, followed by fluctuations. The comparison between the two validates the rationality of classifying predictions based on attrition rates and distribution cycles. These results indicate that the proposed framework provides a statistically grounded and practically effective solution for emergency supply demand forecasting under data-scarce conditions.
Peihua et al. (Tue,) studied this question.