融合改进NBEATSx和时间注意力机制的空气污染预测  

Air pollution prediction by combining improved NBEATSx andtime pattern attention mechanism

作　　者：李杰[1] 王占刚[1] LI Jie;WANG Zhan-gang(School of Information and Communication Engineering,Beijing Information Science&Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学信息与通信工程学院,北京100101

出　　处：《陕西科技大学学报》2024年第5期198-205,共8页Journal of Shaanxi University of Science & Technology

基　　金：国家重点研发计划项目(2018YFC1800203);北京市科技创新服务能力建设-基本科研业务费(市级科研类)项目(PXM2019-014224-000026)。

摘　　要：针对现有空气污染预测存在结构复杂、对多元变量与不同时间步间依赖关系提取不充分和多步预测精度低的问题,引入了β分布和非线性动态控制函数改进星鸦优化算法(INOA),优化NBEATSx模型参数,提高收敛精度;并融合时间模式注意力机制(TPA)为不同时间尺度的多外生变量自适应分配权重,再结合预测因子获取时间模式关系.利用所提模型对北京地区的PM2.5进行预测,与传统模型相比精度提高超过18.45%,为空气污染预测提供了一种新方法.Aiming at the problems of complex structure,inadequate extraction of dependence between multivariate variables and different time steps and low accuracy of multi-step prediction,βdistribution and nonlinear dynamic control function are introduced to improve the Nutcracker optimizer algorithm(INOA),and then INOA is used to optimize the parameters of NBEATSx model to improve the convergence accuracy.The time pattern attention mechanism(TPA)was integrated to assign weights to multiple exogenous variables of different time scales,and the temporal pattern relationship was obtained by combining the predictive factor.The proposed model is used to predict PM2.5 in Beijing,and the accuracy is improved by more than 18.45%compared with the traditional model,which provides a new method for air pollution prediction.

关 键 词：空气污染预测 时间模式注意力机制 星鸦优化算法 神经基扩展分析网络 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]