Forecasting water quality variable using deep learning and weighted averaging ensemble models.

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Computational Ecology

Background:

• Chlorophyll-a (Chl-a) is a key indicator of aquatic ecosystem health, reflecting algal and cyanobacterial populations.
• Accurate forecasting of Chl-a is crucial for effective water quality management and ecological assessment.

Purpose of the Study:

• To assess the predictive performance of four deep learning (DL) models: recurrent neural network (RNN), long short-term memory (LSTM), gated recurrence unit (GRU), and temporal convolutional network (TCN) for Chl-a forecasting.
• To develop ensemble models (EMs) by integrating DL models using genetic algorithm (GA) and non-dominated sorting genetic algorithm II (NSGA-II) to enhance predictive accuracy.
• To evaluate the efficacy of the developed EMs in forecasting Chl-a concentrations.

Main Methods:

• Utilized hourly Chl-a concentration data with lag times up to 6 hours as inputs for forecasting Chl-a (t+1).
• Trained and validated DL and EM models on 70% and 30% of the dataset, respectively, collected from Small Prespa Lake, Greece.
• Employed single-objective GA and multi-objective NSGA-II for developing ensemble models.

Main Results:

• The GRU model demonstrated superior performance among standalone DL models, outperforming RNN, LSTM, and TCN.
• Ensemble models developed using GA and NSGA-II effectively forecasted both low and high Chl-a concentrations.
• The EM-NSGA-II model achieved the highest accuracy, significantly outperforming standalone DL models and the EM-GA across various evaluation metrics, including R-squared.

Conclusions:

• Deep learning models, particularly GRU, show promise for Chl-a forecasting.
• Ensemble modeling with NSGA-II offers a robust approach to improve Chl-a prediction accuracy in aquatic ecosystems.
• The developed EM-NSGA-II provides a highly effective tool for water quality monitoring and management.