基于GCN-GRU-Attention的多站点PM_(2.5)质量浓度预测方法  

Multi-site PM_(2.5) Mass Concentration Prediction Method Based on GCN-GRU-Attention

作　　者：朱振业 唐超礼[2] ZHU Zhenye;TANG Chaoli(School of Computer Science&Engineering,Anhui University of Science and Technology,Huainan 232001,China;School of Electrical&Information Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]安徽理工大学电气与信息工程学院,安徽淮南232001

出　　处：《湖北民族大学学报(自然科学版)》2025年第1期67-72,85,共7页Journal of Hubei Minzu University:Natural Science Edition

基　　金：安徽省研究生创新创业实践项目(2023cxcysj089);安徽理工大学研究生创新基金(2023cx2092)。

摘　　要：为提高大气细颗粒物(particulate matter 2.5,PM_(2.5))质量浓度预测的准确率,并为科学制定空气质量管理策略和污染防治措施提供决策依据,提出了基于图卷积-门控循环单元-注意力机制(graph convolutional network-gated recurrent unit-attention mechanism, GCN-GRU-Attention)组合模型的多站点PM_(2.5)质量浓度预测方法。通过动态构建站点间的时空图捕捉动态相关性,并结合注意力机制增强时空特征的交互性。结果表明,在长期预测中,GCN-GRU-Attention模型相较于对比模型中表现最好的PM_(2.5)图神经网络(PM_(2.5)-graph neural network, PM_(2.5)-GNN)模型,平均绝对误差和均方根误差分别降低了3.2%和1.8%。在短期预测和污染高峰期预测中的效果相较于传统模型和固定图结构模型也均有明显提升。消融实验进一步验证了动态图建模和注意力机制对模型性能提升的有效性。该研究为动态环境下的PM_(2.5)质量浓度预测提供了新思路,并为空气质量管理和公共健康保护提供了可靠依据。To improve the accuracy of particulate matter 2.5(PM_(2.5))mass concentration prediction and provide decision-making support for the formulation of air quality management strategies and the implementation of pollution control measures,the multi-site PM_(2.5) mass concentration prediction method based on a combined model of graph convolutional network-gated recurrent unit-attention mechanism(GCN-GRU-Attention)was proposed.The model dynamically constructed spatiotemporal graphs between monitoring sites to capture dynamic correlations and enhanced the interaction of spatiotemporal features through the attention mechanism.The results showed that,in long-term forecasting,the GCN-GRU-Attention model outperformed the best-performing PM_(2.5)-graph neural network(PM_(2.5)-GNN)model among the comparison models,with reduction of 3.2%in mean absolute error and 1.8%in root mean squared error.In short-term prediction and high pollution peak prediction,significant enhancements were also observed compared to traditional models and fixed graph structure models.These results demonstrated significant improvements over traditional models and fixed graph structure models.The ablation experiments further confirmed the effectiveness of dynamic graph modeling and the attention mechanism in enhancing model performance.This research provided new insights into PM_(2.5) mass concentration prediction in dynamic environments and offered a reliable basis for air quality management and public health protection.

关 键 词：PM_(2.5) 质量浓度预测 动态图 注意力机制 时空特征融合 污染治理 安徽省 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]