Data-driven models for forecasting algal biomass in a large and deep reservoir

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Accurate prediction of algal biomass is vital for drinking water preservation and management.
• Forecasting algal biomass in large, deep reservoirs presents significant challenges.
• High-frequency observational data is essential for developing robust predictive models.

Purpose of the Study:

• To develop and evaluate Long Short-Term Memory (LSTM) models for forecasting algal biomass in a large Chinese reservoir.
• To predict chlorophyll-a concentration (C_Chla) and column-integrated C_Chla (CIC) at various forecasting scales.
• To compare the performance of LSTM models against other machine learning models (MLP, CNN, RNN).

Main Methods:

• Utilized six years of high-frequency (30-min) observational data from Xin'anjiang Reservoir.
• Developed five LSTM-based models: four for C_Chla (1-hr, 3-hr, 6-hr, 24-hr scales) and one for CIC (1-day scale).
• Assessed model performance using root mean square error (RMSE) and compared with MLP, CNN, CNN-LSTM, and RNN models.

Main Results:

• LSTM models accurately predicted C_Chla (RMSE < 1.1 μg/L) and CIC (RMSE < 14.9 μg/L).
• The proposed C_Chla LSTM model outperformed others by 2.6%–9.3% RMSE reduction; the CIC LSTM model showed superior performance with 36.1%–52.8% RMSE reduction.
• Model performance improved with input time length (6-8 times forecasting length) and was better at sites with less biomass variation; water temperature was the key factor.

Conclusions:

• LSTM models provide an effective tool for accurate, large-scale algal biomass forecasting in deep reservoirs.
• These models offer a valuable early warning system for water resource managers to implement preemptive measures against algal blooms.
• Understanding the influence of factors like water temperature is crucial for refining predictive accuracy.