Predicting ammonia nitrogen in surface water by a new attention-based deep learning hybrid model

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Artificial Intelligence in Environmental Management

Background:

• Ammonia nitrogen (NH3-N) is a critical factor in surface water ecosystem health, influencing cyanobacterial blooms.
• Traditional predictive models struggle with the complex, nonlinear relationships between NH3-N and environmental parameters.
• Accurate NH3-N forecasting is essential for effective water resource management and ecosystem protection.

Purpose of the Study:

• To develop a superior predictive model for ammonia nitrogen (NH3-N) concentrations in surface waters.
• To address the limitations of traditional models in capturing complex nonlinear environmental relationships.
• To enhance the accuracy and reliability of water quality forecasting for early warning systems.

Main Methods:

• Proposed a novel hybrid model: Boundary Corrected Maximal Overlap Discrete Wavelet Transform-Dual-Stage Attention-Long Short-Term Memory (BC-MODWT-DA-LSTM).
• Utilized BC-MODWT for data decomposition to identify trends and swings in environmental feature series.
• Integrated a Dual-Stage Attention (DA) mechanism to enable the LSTM network to focus selectively on input data.

Main Results:

• The BC-MODWT-DA-LSTM model demonstrated superior performance compared to other models, exhibiting lower average prediction errors.
• Achieved high predictive accuracy with NASH Sutcliffe efficiency coefficient (NSE) values above 0.900 for lead times up to 7 days.
• The area under the receiver operating characteristic (ROC) curve reached 0.992, indicating excellent classification performance.
• Showcased high prediction accuracy at peak pollution events, enabling early warning capabilities for sudden NH3-N surges.

Conclusions:

• The developed BC-MODWT-DA-LSTM hybrid model offers a promising approach for accurate water quality prediction.
• Enhancing LSTM models without increased topological complexity is an effective strategy for improving forecasting accuracy.
• The model's capability for early warning of high NH3-N pollution is crucial for proactive environmental management.