改进DTW下界约束的Granger多元时序LSTM预测模型  

GRANGER MULTIVARIATE TIME SERIES LSTM PREDICTION MODEL WITH IMPROVED DTW LOWER BOUND CONSTRAINT

作　　者：许凤魁 孙士保 贾少勇 王静 Xu Fengkui;Sun Shibao;Jia Shaoyong;Wang Jing(College of Information Engineering,Henan University of Science and Technology,Luoyang 471023,Henan,China)

机构地区：[1]河南科技大学信息工程学院,河南洛阳471023

出　　处：《计算机应用与软件》2024年第5期233-239,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(51474095);河南省重点攻关项目(152102210277);河南省高校科技创新团队支持计划项目(17IRTSTHN010);河南科技大学科技创新团队项目(2015XTD011);河南科技大学重大产学研合作培育基金项目(2015ZDCXY03)。

摘　　要：多元时序的因果预测研究是探讨复杂网络响应关系的热点问题。提出一种通过DTW的下界约束组合并改进的层级过滤器,与格兰杰因果验证法相结合验证其因果统计量,挖掘出有效信息实现有效降维,进一步输入LSTM预测模型进行因果预测。仿真实验利用开源的空气质量数据集进行定量和定性对比验证,该方法的损失函数训练曲线和测试曲线有较好的拟合度,表明该因果预测法是可行且有效的。The research on the causal prediction of multivariate time series is a hot issue to explore the relationship between complex networkdriven responses.This paper proposes a hierarchical filter method that combines and improves the lower bound constraints of dynamic time warping(DTW).This method was combined with Granger causality to verify the causal statistics,so that it dug out effective value information to achieve effective dimensionality reduction.And it was inputted into LSTM prediction model to make causal predictions.The simulation experiment used opensource air quality time series data sets for quantitative and qualitative comparison and verification.It is found that the training curve and the test curve in its loss function curve have a better fit,which shows that the causal prediction method is feasible and effective.

关 键 词：动态时间弯曲 长短时记忆网络 格兰杰因果关系 层级过滤器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]