摘要
针对在臭氧质量浓度预测中时空特征建模不足与时间动态依赖捕捉不充分的问题,提出基于多尺度图空间特征提取-时间注意力增强门控循环单元(multi-scale graph spatial feature extraction-temporal attention enhanced gated recurrent unit,MSGSFE-TAGRU)预测模型。该模型由MSGSFE与TAGRU模块构成。MSGSFE模块通过结合图注意力网络(graph attention network,GAT)与图卷积网络(graph convolutional network,GCN)构建多尺度图结构,实现对局部与全局空间特征的有效提取;TAGRU模块通过引入时间注意力机制动态聚焦关键历史时刻,提升门控循环单元(gated recurrent unit,GRU)对时序依赖的建模能力。结果表明,MSGSFE-TAGRU模型在短期预测中相较于表现最优的图神经网络-GRU(graph neural network-GRU,GNN-GRU)模型,均方根误差和平均绝对误差分别降低了5.22%和5.76%;在中长期预测中,该模型亦表现出更优的精度和稳定性,验证了其卓越的时空特征建模能力与泛化性能。该研究为臭氧质量浓度的高精度预测和空气质量管理提供了新的方法框架。
To address the limitations in spatiotemporal feature modeling and the insufficient capture of temporal dependencies in ozone mass concentration prediction,a multi-scale graph spatial feature extraction-temporal attention enhanced gated recurrent unit(MSGSFE-TAGRU)prediction model was proposed.The model consisted of two main modules:MSGSFE and TAGRU modules.In the MSGSFE module,multi-scale graph structures were constructed by integrating graph attention network(GAT)and graph convolutional network(GCN)to effectively extract both local and global spatial features.In the TAGRU module,a temporal attention mechanism was introduced to dynamically focus on key historical time steps,which enhanced the capability of the gated recurrent unit(GRU)in modeling for temporal dependencies.The results demonstrated that,in short-term prediction,the MSGSFE-TAGRU model reduced root mean square error and mean absolute error by 5.22%and 5.76%,respectively,compared to the best-performing graph neural network-GRU(GNN-GRU)model.In medium-and long-term prediction,the proposed model continued to exhibit superior accuracy and stability,validating its effectiveness in spatiotemporal feature modeling and generalization.This study provided a novel methodological framework for high-precision prediction of ozone mass concentration and air quality management.
作者
甄点点
唐超礼
朱振业
ZHEN Diandian;TANG Chaoli;ZHU Zhenye(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China;School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan 232001,China)
出处
《湖北民族大学学报(自然科学版)》
2025年第3期425-430,共6页
Journal of Hubei Minzu University:Natural Science Edition
基金
安徽理工大学研究生创新基金项目(2024cx2114,2024cx2076)。
关键词
臭氧质量浓度预测
多尺度图卷积
动态时间注意力机制
时空预测模型
北京市
ozone mass concentration prediction
multi-scale graph convolution
dynamic temporal attention mechanism
spatiotemporal prediction model
Beijing City
