Hybrid decision tree-based machine learning models for short-term water quality prediction.

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Water resources are vital for human life, economic development, health, and the environment.
• Accurate water quality prediction is essential for effective water management and pollution control.
• The Tualatin River Basin, a globally polluted river system, serves as a case study.

Purpose of the Study:

• To develop and evaluate novel hybrid machine learning models for short-term water quality prediction.
• To enhance prediction accuracy by integrating advanced data denoising techniques with decision tree-based models.
• To assess the performance and stability of the proposed models against conventional methods.

Main Methods:

• Two hybrid models were developed: Random Forest (RF) with Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Extreme Gradient Boosting (XGBoost) with CEEMDAN.
• Hourly data for six water quality indicators (water temperature, dissolved oxygen, pH, specific conductance, turbidity, fluorescent dissolved organic matter) were collected over a 79-day period.
• Model performance was evaluated using six error metrics, including Mean Absolute Percentage Error (MAPE), and compared against four benchmark models.

Main Results:

• CEEMDAN-RF excelled in predicting temperature (MAPE 0.69%), dissolved oxygen (MAPE 1.05%), and specific conductance (MAPE 0.90%).
• CEEMDAN-XGBoost demonstrated superior performance in predicting pH (MAPE 0.27%), turbidity (MAPE 14.94%), and fluorescent dissolved organic matter (MAPE 1.59%).
• Both CEEMDAN-RF and CEEMDAN-XGBoost achieved the lowest average MAPEs (3.90% and 3.71%, respectively), indicating the best overall prediction accuracy and stability.

Conclusions:

• The hybrid CEEMDAN-RF and CEEMDAN-XGBoost models offer significant improvements in short-term water quality prediction accuracy.
• These advanced models provide more stable and reliable predictions compared to conventional methods.
• The findings support the application of these hybrid models for enhanced water resource management and pollution control strategies.