科技通报

地震灾害等突发事件对人类社会造成严重威胁，多种因素的不确定性导致应急救援生活保障物资需求量具有强烈的不确定性，准确预测应急物资需求量一直是亟需解决的难题。本文以震后各类生活保障物资为研究对象，融合多种机器学习模型的优点，构建了面向物资需求预测的异构集成框架，设计了“交叉验证+网格搜索”的双层优化机制，利用实际数据进行案例分析，并将本方法与9类基线模型进行对比，通过消融实验揭示了初级学习器的动态交互机制。结果表明：相较于基线模型，本模型在平均绝对百分比误差（mean absolute percentage error,MAPE）方面降低了5.65%～87.28%，预测结果准确性较高，能够为应急救援提供方法支撑和决策依据。

Earthquakes and similar emergencies severely threaten human society. Emergency relief supplies are critical for disaster response, and their supply-demand matching directly affects rescue effectiveness. The demand for emergency living supplies faces high uncertainty due to multiple factors, making accurate prediction a persistent challenge. Current engineering practices rely on subjective expert experience, which struggles to meet real-world emergency response needs. Existing research primarily uses single prediction models with limited accuracy and reliability. To address these issues, this study develops a heterogeneous ensemble framework for predicting postearthquake living supplies demand. The framework integrates historical data,disaster types, affected regions,and other key factors. It combines advantages of multiple machine learning models and implements a two-tier optimization mechanism with cross-validation and grid search. The method is tested using magnitude ≥5 earthquake data from Yunnan Province since 2007. Comparisons with nine baseline models—including random forest(RF), gradient boosting(GB), deep convolutional neural network(DCNN), K-nearest neighbor(KNN), support vector machine(SVM), and back propagation neural network(BPNN) —show significant improvements. The proposed model reduces mean absolute percentage error(MAPE) by 5.65% – 87.28% compared to baseline models. Ablation experiments further analyze dynamic interactions among base learners. The results demonstrate high prediction accuracy, providing actionable decision-making support for emergency rescue operations.