首页|基于降方差采样策略的随机重构法

基于降方差采样策略的随机重构法

扫码查看
在随机重构法的采样过程中,每个结构特征需要的采样数量是不相等且变化的.为了将拉丁超立方采样用于降低随机重构模型的方差,基于随机重构法采样过程的特征和拉丁超立方采样原理,提出了适用于随机重构法的新型拉丁超立方采样方法,探究了在多种分子数量设定情况下应用该方法对随机重构模型的方差和精度的影响.结果表明,应用该方法能够显著降低随机重构模型的方差,提高模型的精度,在分子数量为1000~50000范围内,新模型的标准差相较传统模型降低了71.36%~74.53%,目标函数值降低了1.69%~13.82%.综合模型精度和模拟过程的运算开销,选择4000~6000作为新模型最优的分子数量设定.
Variance reduction sampling strategy-based stochastic reconstruction method
In the sampling procedure of the stochastic reconstruction method,the number of samples for each structural attribute is unequal and varying.To use Latin hypercube sampling to reduce the variance of the random reconstruction model,based on the characteristics of the sampling process of the stochastic reconstruction method and the principle of Latin hypercube sampling,a new Latin hypercube sampling method suitable for the stochastic reconstruction method was proposed.In the novel Latin hypercube sampling,the sampling number for each structural attribute is determined by the values of the preorder structural attribute.A Latin hypercube sampling-based stochastic reconstruction model for a vacuum gas oil sample was developed.The determination of the sampling number for each structural attribute was introduced in detail with the building diagram.Multiple cases with different predefined molecular numbers were designed to investigate the effects of the novel Latin hypercube sampling on the variance and accuracy of the stochastic reconstruction model.The results showed that the novel Latin hypercube sampling method can significantly reduce the variance and objective function value of the model.As the molecular number ranging from 1000 to 50000,the standard deviations of the new model are 71.36%—74.53%lower than the traditional model,and the objective function values are 1.69%—13.82%lower than the traditional model.Based on the model accuracy and the computational cost of the simulation process,4000-6000 was selected as the optimal molecule numbers for the new model.By comparing the values of objective function,bulk properties and mass fraction distributions in saturates and aromatics,it was found that the performance of the new model when the molecular number is 2000 is consistent with the performance of the traditional model when the molecular number is 10000,and the computing time of the new model is only 22.54%of that of the traditional model.

stochastic reconstruction methodnovel Latin hypercube samplingvariancesimulationMonte Carlo simulation

赵光耀、杨明磊、钱锋

展开 >

华东理工大学能源化工过程智能制造教育部重点实验室,上海 200237

华东理工大学过程系统工程教育部工程研究中心,上海 200237

随机重构法 新型拉丁超立方采样 方差 模拟 Monte Carlo模拟

国家重点研发计划国家自然科学基金重大项目国家自然科学基金面上项目国家自然科学基金青年科学基金中央高校基本科研业务费专项

2022YFB3305900622935016237315362203173

2024

化工学报
中国化工学会 化学工业出版社

化工学报

CSTPCD北大核心
影响因子:1.26
ISSN:0438-1157
年,卷(期):2024.75(5)