Stable Lagged Dependency Graphs and Anomaly-Driven Adaptation for Dissolved Oxygen Forecasting

摘要

Accurate multi-horizon dissolved oxygen (DO) forecasting is crucial for environmental health assessment and early warning. However, the impacts of environmental disturbances on DO make the forecasting challenging because DO dynamics combine slowly varying background evolution with delayed and event-dependent environmental interactions. Existing temporal models generally treat environmental disturbances as ordinary inputs, while fixed or fully adaptive graphs do not clearly separate persistent lagged dependencies from transient event responses. To resolve these issues, we propose SLDGA – DO, a Stable Lagged Dependency Graph Adaptation framework for direct multi-horizon DO forecasting. SLDGA-DO constructs station-specific stable dependency graphs through multi-scale detrending, fragment-level subsampling, and PCMCI+-based lagged dependency discovery. During forecasting, explicit thermal and runoff states reweight dependencies within the stable graph and augment the node representations with transient interactions. A graph--temporal encoder models short-term fluctuations, while an independent trend branch preserves the slowly varying background trajectory. Experiments on five Chesapeake Bay stations show that SLDGA-DO achieves the best R^2 across all forecasting horizons and reduces long-horizon MAE by 40.0%--43.3% over the strongest baselines. It also obtains the best results under thermal and runoff events, including a 38.6% MAE reduction for long-horizon runoff forecasting. Ablation results confirm the contributions of decomposition, graph--temporal encoding, and trend modeling.