The fast and accurate estimation of makespan is essential for the determination of the delivery date and the sustainable development of the enterprise. In this paper, a high-quality training dataset is constructed and an adaptive ensemble model is proposed to achieve fast and accurate makespan estimation. First, both the logistics features extracted by the Pearson correlation coefficient and the new meaningful nonlinear combination features dug out by gene expression programming are first involved in this paper for constructing a high-quality dataset. Secondly, an improved clustering with elbow criterion and a resampling operation are applied simultaneously to generate representative subsets; and correspondingly, several back propagation neural network (BPNN) with the architecture optimized by genetic algorithm are trained by these subsets respectively to generate effective diverse learners; and then, a K-nearest neighbor based dynamic weight combination strategy which is sensitive to current testing sample is proposed to make full use of the learner’s positive effects and avoid its negative effects. Finally, the results of effective experiments prove that both the newly involved features and the improvements in the proposed ensemble are effective. In addition, comparison experiments confirm that the proposed enhanced ensemble of BPNNs outperforms significantly the prevailing approaches, including single, ensemble and hybrid models. And hence, the proposed model can be utilized as a convenient and reliable tool to support customer order acceptance.

快速准确地估计制造期限对于确定交货日期和企业的可持续发展至关重要。本文构建了一个高质量的训练数据集，并提出了一种自适应集成模型来实现快速准确的估计工期。首先，本文首先涉及利用Pearson相关系数提取的物流特征和通过基因表达编程挖掘出的新的有意义的非线性组合特征，以构建高质量的数据集。其次，同时应用了经过改进的基于肘部准则的聚类和重采样操作，以生成代表性子集。相应地，这些子集分别训练了几个具有遗传算法优化结构的反向传播神经网络（BPNN），以产生有效的多样化学习者。然后，为了充分利用学习者的积极影响，避免学习者的消极影响，提出了一种对当前测试样本敏感的基于K近邻的动态权重组合策略。最后，有效实验的结果证明，新涉及的功能和所提出的合奏中的改进都是有效的。此外，比较实验证实，所提出的增强型BPNN集成明显优于主流方法，包括单模型，集成模型和混合模型。因此，所提出的模型可以用作方便可靠的工具来支持客户订单接受。