Decomposition prediction and optimal ensemble strategy improve river dissolved oxygen prediction accuracy.

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Accurate dissolved oxygen (DO) prediction is vital for managing riverine ecosystems.
• Existing hybrid models for nonlinear DO prediction often fall short.

Purpose of the Study:

• To propose a frequency division prediction framework using optimal ensemble methods for enhanced DO prediction in rivers.
• To address the limitations of current hybrid models in capturing nonlinear DO dynamics.

Main Methods:

• Utilized Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose DO time series into multiple components.
• Developed Long Short-Term Memory (LSTM), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP) models for independent component prediction.
• Implemented a constrained grid search algorithm for dynamic model ensemble optimization, minimizing Mean Absolute Error (MAE).

Main Results:

• The proposed integrated model demonstrated significant improvements over ensemble models, with MAE reductions of 18.6-35.5% and Root Mean Square Error (RMSE) reductions of 22.1-22.8%.
• Determination coefficients (R2) reached 0.954 and 0.972, indicating high prediction accuracy.
• A notable 27.2-81.4% reduction in 3-day prediction MAE error accumulation was observed compared to mixed models.

Conclusions:

• The frequency division prediction framework effectively integrates multi-component DO series predictions.
• This approach offers an extensible technical solution for intelligent river basin management and improved water quality monitoring.