Bridging prediction and causal attribution in large-scale river water-quality networks

Water-quality forecasting and attribution at large spatial scales remain challenging because observations are sparse and heterogeneous, monitoring networks are nonstationary, and river systems impose strong directional and time-delayed connectivity. These constraints are further complicated by abrupt pollution shocks that cannot be explained by routine upstream–downstream propagation, yet are critical for risk-aware water management. Consequently, existing approaches often face a trade-off between predictive accuracy and interpretability: data-driven models capture complex spatiotemporal patterns but provide limited insight into underlying drivers, while causal analyses offer mechanistic understanding but are difficult to operationalize for real-time, multi-step forecasting across large monitoring networks.

Here we develop a physically constrained, topology-aware causal forecasting framework that unifies large-scale water-quality prediction with driver attribution in river networks. The framework explicitly represents three defining characteristics of fluvial systems: unidirectional upstream-to-downstream transport, travel-time-dependent propagation delays, and dynamic monitoring configurations in which stations appear or disappear over time. By embedding physical flow constraints into a data-driven causal representation, the framework jointly learns evolving spatiotemporal dependencies while remaining robust to extreme data sparsity and uneven sampling typical of water-quality observations.

We apply the framework to China’s national surface-water monitoring network, comprising more than 1,900 stations and multiple water-quality indicators, together with hydro-meteorological covariates. The framework achieves strong multi-step predictive skill across the full network under realistic data gaps, while providing an interpretable decomposition of dynamics into local persistence, upstream propagation, and externally driven disturbances. This decomposition enables real-time identification of dominant drivers of water-quality change at individual stations, distinguishes systemic trends from short-lived pollution shocks, and localizes influential upstream contributors consistent with river-network connectivity. Beyond forecasting, the framework supports event-oriented causal diagnostics and prioritization of high-information monitoring locations, helping to optimize sampling strategies and enhance early-warning capability under limited monitoring resources.

Our results demonstrate that physically constrained causal forecasting can bridge the long-standing divide between prediction and explanation in water-quality modelling at scale. Crucially, the framework remains operational under nonstationary station availability, enabling consistent forecasting and attribution as monitoring configurations evolve. By integrating topology-aware learning with interpretable attribution, the proposed framework establishes a coherent pathway from forecasting to diagnosis and source localization, supporting proactive, data-informed water-quality management and rapid response to pollution events in complex river networks.