Two-stage meta-ensembling machine learning model for enhanced water quality forecasting

Abstract

Accurate short-term forecasting of water quality variables (WQVs) such as dissolved oxygen (DO) and chlorophyll-a (Chl-a) is crucial for the effective management of aquatic resources. This study introduces a robust two-stage optimization-ensembling framework that integrates the Grey Wolf Optimizer (GWO) and the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to enhance the forecasting capabilities of machine learning (ML) models. Focusing on Small Prespa Lake, Greece, we implemented an array of diverse ML techniques, including eXtreme Gradient Boosting (XGB), Gradient Boosting Regressor (GBR), Light Gradient-Boosting Machine (LightGBM), and Multilayer Perceptron (MLP). These models were fine-tuned using GWO to optimize their performance over critical water quality parameters predicted six hours in advance. Our methodology employed rigorous data preprocessing techniques such as lag time feature engineering and principal component analysis (PCA) to manage the high dimensionality of the dataset. Optimal lag times ranging from 6 to 24 h were evaluated, with a 24-hour lag proving most effective in leveraging historical data to enhance forecasting accuracy. The GWO not only facilitated hyperparameter tuning but also demonstrated a significant 7.6% improvement in the Kling-Gupta Efficiency (KGE) over conventional randomized search methods. Subsequently, the NSGA-II was utilized for multi-objective optimization to construct powerful model ensembles that outperformed the individual GWO-optimized models, showing up to a 7% increase in KGE. When compared to a standard genetic algorithm-based ensemble, the NSGA-II ensemble displayed enhanced effectiveness in balancing solution quality. This innovative approach not only establishes a new standard in water quality forecasting but also contributes substantially to proactive environmental monitoring and management strategies.