Enhanced interpretable data driven-based model for dissolved oxygen prediction under multiple environmental factor influences

Rationale: Dissolved oxygen (DO) is a critical indicator of water quality, but accurately predicting DO fluctuations under the combined influence of various environmental factors remains challenging. Methodology: This study presents an intelligent prediction framework that integrates interpretable machine learning methods with stacking generalization techniques to forecast dynamic changes in water quality effectively. Using daily data (2020-2022) from seven stations in the Danjiangkou Reservoir, China, environmental factors were categorized into water quality, streamflow, meteorology, and air pollution. A stacked model, combining SVM, MLP, RF, and KNN base learners with a LightGBM meta-learner, was constructed and interpreted using SHAP analysis. The contribution of each feature is accurately determined, and the mode of action of the main factors and the key threshold points are identified. Results: The stacked model provides efficient and accurate DO forecasts, outperforming single base models with an average improvement in R² of 5.22% on the test set, and achieving a 2.29% increase in R² and a 4.81% reduction in RMSE compared to the unstacked LightGBM model. Global interpretation reveals that the water quality parameter category has the most significant impact (up to 70% of total feature importance), with water temperature being the most crucial factor. Other key contributors include conductivity, ammonia turbidity, runoff volume, atmospheric pressure , and sulfur dioxide. Discussion: The proposed framework effectively captures the synergistic effects of multiple environmental factors on DO. It offers a robust method for forecasting their dynamics, demonstrating the value of categorized environmental analysis for water quality management.