Sensor-Aware Graph Convolutional and LSTM Model for Reliable Water Quality Forecasting

This paper addresses the critical challenge of unreliable water quality forecasting caused by heterogeneous sensor data, measurement uncertainties, and complex spatio-temporal dynamics in aquatic environments. Traditional forecasting methods struggle with sensor drift, varied error margins, and irregular sampling frequencies, leading to inaccurate predictions that hamper effective water resource management and pollution control. To overcome these limitations, we present a novel hybrid deep learning framework that combines graph convolutional networks (GCNs) with long short-term memory (LSTM) architectures. By explicitly embedding sensor-specific uncertainties into both spatial and temporal modeling, the proposed method robustly handles heterogeneous sensor data from diverse monitoring stations. The GCN component constructs an adjacency matrix from geodesic or Euclidean distances, capturing localized interactions critical for detecting pollution hotspots or resource strains. The LSTM module processes extended historical sequences of water quality indicators to capture seasonal, diurnal, and episodic fluctuations. Empirical tests on multi-season datasets demonstrate the framework’s superior predictive performance compared to traditional methods, with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values up to 0.91 for temperature predictions and reliable confidence intervals through ensemble-based uncertainty estimates.