Wan Hang, Xiang Long, Cai Yanpeng, Xie Yulei, Xu Rui
Research Centre of Ecology & Environment for Coastal Area and Deep Sea, Southern Marine Science and Engineering Guangdong Laboratory (Guangzhou), Guangzhou 511458, China.
School of Computer Science and Information Security, Guilin University of Electronic Technology, Guilin 541004, China.
Water Res. 2025 Mar 26;281:123561. doi: 10.1016/j.watres.2025.123561.
Deep learning methods have demonstrated strong capabilities in capturing nonlinear relationships for water quality prediction, yet existing studies predominantly focus on individual monitoring sites while neglecting pollutant spatial dynamics. To address this limitation, a Spatio-Temporal Feature Graph Neural Network (STF-GNN) was proposed, which integrated graph convolutional networks (GCN), gated recurrent units (GRU), and self-attention mechanisms to explicitly model multi-scale spatiotemporal dependencies among distributed monitoring stations. By representing stations as graph nodes with adjacency relationships, STF-GNN could simultaneously extract spatial topological features and temporal evolution patterns from multivariate time series data. Experimental results demonstrated superior performance in dissolved oxygen (DO) and total nitrogen (TN) prediction, achieving RMSE values of 0.233 (DO) and 0.033 (TN), outperforming baseline models by 36.54-161.47 % in accuracy. Cross-basin validations revealed robust generalization capabilities of the established model, maintaining maximum relative errors below 0.639 (DO) and 0.606 (TN) without site-specific customization. Notably, the model achieved 88 % peak-valley synchronization at untrained station, demonstrating strong anti-interference ability against unseen environmental variations. Ablation studies confirmed the necessity of both spatial and temporal modules, with their omission causing significant accuracy declines (12.07-19.25 %). These findings highlighted the critical roles of both spatial and temporal feature extraction in improving predictive performance of the model. The work can provide a theoretically grounded framework for spatially-aware water quality prediction, supporting enhanced environmental monitoring strategies.
深度学习方法在捕捉水质预测中的非线性关系方面已展现出强大能力，但现有研究主要聚焦于单个监测站点，而忽视了污染物的空间动态。为解决这一局限性，提出了一种时空特征图神经网络（STF-GNN），它集成了图卷积网络（GCN）、门控循环单元（GRU）和自注意力机制，以明确建模分布式监测站之间的多尺度时空依赖性。通过将站点表示为具有邻接关系的图节点，STF-GNN能够同时从多变量时间序列数据中提取空间拓扑特征和时间演化模式。实验结果表明，该方法在溶解氧（DO）和总氮（TN）预测方面具有卓越性能，RMSE值分别为0.233（DO）和0.033（TN），在准确率上比基线模型高出36.54-161.47%。跨流域验证揭示了所建立模型强大的泛化能力，在无需特定站点定制的情况下，最大相对误差保持在0.639（DO）和0.606（TN）以下。值得注意的是，该模型在未训练站点实现了88%的峰谷同步，表明对未见过的环境变化具有强大的抗干扰能力。消融研究证实了空间和时间模块的必要性，省略它们会导致显著的准确率下降（12.07-19.25%）。这些发现突出了空间和时间特征提取在提高模型预测性能方面的关键作用。这项工作可为具有空间感知能力的水质预测提供一个理论基础框架，支持强化环境监测策略。
