A multivariate water quality forecasting model with dynamic variable selection and dissolved oxygen...

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Accurate water quality prediction is crucial for ecological health and resource management.
• High-frequency monitoring data often suffer from noise and complex variations, challenging data-driven models.
• Existing models struggle with unstable or physically inconsistent predictions during extreme conditions.

Purpose of the Study:

• To develop a unified framework for short-term, high-frequency water quality prediction, specifically for dissolved oxygen.
• To improve the accuracy, stability, and physical credibility of water quality predictions.
• To address limitations of purely data-driven models in handling noisy and complex environmental data.

Main Methods:

• Implemented a learnable multi-scale Savitzky-Golay smoothing fusion (LMSG) for noise reduction and feature preservation.
• Introduced a dynamic variable selection mechanism to account for time-varying influences on dissolved oxygen.
• Utilized Mamba-based temporal encoding for sequence dependency modeling and a Weiss-based soft physical-consistency constraint for output regularization.

Main Results:

• The proposed framework demonstrated competitive or superior performance compared to baseline models across three real-world datasets.
• Effectiveness and interpretability were validated through ablation studies, variable-weight visualization, and sensitivity analysis.
• The method successfully suppressed noise while preserving key data variations and ensured physical plausibility of predictions.

Conclusions:

• The integrated framework offers a robust solution for short-term, high-frequency water quality prediction.
• It effectively balances prediction accuracy, model stability, interpretability, and physical consistency.
• This approach provides a valuable tool for water quality assessment and environmental management.