| 本文已被：浏览 1676次 下载 1377次 |
码上扫一扫！ |
|
|
| TaGNN: 基于趋势感知图神经网络的耦合噪声水质预测 |
|
孙建明1,2, 徐宇扬1, 仝硕3, 应豪超3, 张啸4, 庄福振5, 吴健3
|
|
1.浙江大学 计算机科学与技术学院, 浙江 杭州 310058;2.苏州市伏泰信息科技股份有限公司, 江苏 苏州 215124;3.浙江大学 公共卫生学院, 浙江 杭州 310058;4.山东大学 计算机科学与技术学院, 山东 青岛 266237;5.北京航空航天大学 人工智能学院, 北京 100191
|
|
| 摘要: |
| 未来水质预测任务根据各个观测节点的历史水质数据及相应的拓扑位置关系, 预测未来水质变化情况, 是图神经网络辅助环境保护的重要任务之一. 然而, 采集数据的数值和节点间的拓扑结构均存在噪声, 且噪声存在耦合现象. 同时, 污染物流向的不同导致数值噪声与结构噪声的耦合现象更为复杂难以解耦. 因此, 提出了一种基于趋势感知图神经网络的耦合噪声水质预测方法: 1)利用历史水质数据趋势特征, 挖掘原始水质数据指标的局部相互关系, 构建多种潜在的水系拓扑结构, 分离结构噪声; 2)利用构建的邻接矩阵与原始数据挖掘时空特征, 分离数值噪声. 最终, 根据潜在结构构建前后图网络节点表征一致性聚合水质预测结果. 实验结果表明, 所提方法在真实数据集上比当前最优方法表现更好, 且给出的水系潜在拓扑结构符合真实情况. 代码和数据可在GitHub上访问: https://github.com/aTongs1/TaGNN. |
| 关键词: 水质预测 图神经网络 时空数据 结构噪声 图噪声学习 |
| DOI：10.13328/j.cnki.jos.007381 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金 (62106218); 浙江省自然科学基金 (LY23F020019) |
|
| TaGNN: Tendency-aware Graph Neural Network for Water Quality Prediction with Coupled Noise |
|
SUN Jian-Ming1,2, XU Yu-Yang1, TONG Shuo3, YING Hao-Chao3, ZHANG Xiao4, ZHUANG Fu-Zhen5, WU Jian3
|
|
1.College of Computer Science and Technology, Zhejiang University, Hangzhou 310058, China;2.Suzhou Vortexinfo Technology Co. Ltd., Suzhou 215124, China;3.School of Public Health, Zhejiang University, Hangzhou 310058, China;4.School of Computer Science and Technology, Shandong University, Qingdao 266237, China;5.School of Artificial Intelligence, Beihang University, Beijing 100191, China
|
| Abstract: |
| The prediction of future water quality, which involves leveraging historical water quality data from various observation nodes and their corresponding topological relationships, is recognized as a critical application of graph neural networks in environmental protection. This task is complicated by the presence of noise within both the collected numerical data and the inter-node topological structures, compounded by a coupling phenomenon. The varying directions of pollutant flow intensify the complexity of coupling between numerical and structural noise. To address these challenges, a novel tendency-aware graph neural network is proposed for water quality prediction with coupled noise. First, historical water quality trend features are used to uncover local interdependencies among raw water quality indicators, enabling the construction of multiple potential hydrological topological structures and the disentanglement of structural noise. Second, spatio-temporal features are extracted from the constructed adjacency matrices and original data to separate numerical noise. Finally, water quality predictions are obtained by aggregating coherent node representations derived from the inferred latent structures across pre- and post-structure construction phases. Experimental results demonstrate that the proposed method outperforms state-of-the-art models on real-world datasets and generates potential hydrological topological structures that closely align with actual observations. The code and data are publicly available on GitHub: https://github.com/aTongs1/TaGNN. |
| Key words: water quality prediction (WQP) graph neural network (GNN) spatio-temporal data structural noise graph noise learning |
