Attention-based self-learning graph neural networks for robust multi-node water quality prediction
TL;DR
An Attention-based Multivariate Time Series Graph Neural Network (AMTGNN) is proposed for multi-node water quality prediction, which dynamically learns the correlations between different locations to jointly model temporal and spatial dependencies for improved prediction accuracy and robustness.
Abstract
Water quality prediction plays a critical role in water supply systems (WSSs) management such as optimizing chlorination, identifying anomalies, and tracing pollutants. In contrast to traditional single-objective prediction, multi-node prediction holds increasing importance given the rising monitoring capabilities today. Despite several existing studies on multi-node water quality prediction, how to leverage the spatiotemporal relationships among monitoring point data to enhance prediction accuracy and stability, and how to ensure model robustness under abnormal conditions such as missing data are still problems that have not yet been fully addressed. This study proposes an Attention-based Multivariate Time Series Graph Neural Network (AMTGNN) for multi-node water quality prediction, which dynamically learns the correlations between different locations to jointly model temporal and spatial dependencies for improved prediction accuracy and robustness. The model is tested on a real-world urban-rural WSS. A comparative analysis with other state-of-the-art models is conducted, and the impact of different inputs and the robustness under missing data scenarios are also investigated. Experimental results show that AMTGNN achieves an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}$$\end{document} of 0.92 when predicting residual chlorine variation 3 h ahead, and maintains a Mean Absolute Percentage Error (MAPE) below 5.5% with an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}$$\end{document} above 0.85 even under 90% missing data, demonstrating robustness under realistic and imperfect WSS conditions.