Time-series forecasting faces a major challenge when input data is missing. Additionally, standard multi-site water quality models often fail to capture the spatial connections among monitoring stations. This study proposes a GAT-enhanced LSTM model (GAT-LSTM) that integrates Graph Attention Networks (GAT) with Long Short-Term Memory (LSTM) to enhance prediction robustness under data incompleteness. We established a systematic evaluation framework, using MAE, MAPE, and R 2 as the metrics for assessing predictive performance. In addition, we defined a Comprehensive Robustness Index (CRI) to evaluate model performance under three scenarios: spatial (missing stations), temporal (missing time steps), and random (missing indicators). Using real-world data from 13 monitoring stations in Pearl River, the third largest river in China, we compared GAT-LSTM against a standalone LSTM. Results show that the two models achieved comparable accuracy when data were complete; however, across all missing-data scenarios, GAT-LSTM consistently demonstrated superior robustness, exhibiting 1.3–1.8 times greater tolerance to data loss than the conventional LSTM. The GAT component became critical when spatial data is missing. The performance gap was most pronounced when key monitoring stations were removed first: GAT-LSTM maintained high stability (CRI: 0.98), whereas the standalone LSTM experienced a sharp decline (CRI: 0.5). These findings confirm that incorporating the GAT architecture provides powerful compensatory capability for incomplete spatial data, rendering GAT-LSTM significantly more resilient in real-world water quality prediction tasks. When monitoring networks suffer from inconsistent spatial coverage, GAT transitions from an optional enhancement to an essential core component. • A GAT-LSTM model leverages spatial topology for robust water quality prediction. • A novel framework quantifies model robustness against diverse data loss scenarios. • GAT-LSTM is 1.3–1.8 times more robust to data loss than standalone LSTM. • River spatial topology becomes essential when critical monitoring sites are lost.
Wang et al. (Sun,) studied this question.