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

Accurate water quality prediction is important for water quality assessment, aquatic ecosystem protection, and water-environment management. However, high-frequency monitoring data are often affected by strong noise, multi-scale variations, and time-varying driving factors, making purely data-driven models prone to unstable or physically inconsistent predictions under abrupt or extreme conditions. To address these issues, this study presents a unified short-term high-frequency water quality prediction framework that integrates learnable multi-scale preprocessing, dynamic variable selection, Mamba-based temporal encoding, and a Weiss-based soft physical-consistency constraint for dissolved oxygen. First, a learnable multi-scale Savitzky-Golay smoothing fusion method (LMSG) is developed to suppress high-frequency noise while preserving key peaks and local variations. Second, a dynamic variable selection mechanism is introduced to characterize the time-varying effects of multivariate water quality parameters on dissolved oxygen, improving predictive stability and interpretability. Third, Mamba-based temporal encoding is used to model dependency patterns within high-frequency input sequences, while the Weiss-based dissolved oxygen physical-consistency term serves as a soft output-level regularizer to improve the internal consistency of predicted dissolved oxygen representations and penalize evident supersaturation. Experimental results on three real-world water quality monitoring datasets indicate that the proposed method achieves competitive and generally better performance than the selected representative baseline models under the unified experimental setting of this study. Ablation experiments, variable-weight visualization, and perturbation sensitivity analysis further support the effectiveness and interpretability of the proposed framework. Overall, the presented approach provides a useful framework for short-term high-frequency water quality prediction by balancing accuracy, stability, interpretability, and physical credibility within the scope of the investigated datasets.