A hybrid deep learning approach to predict hourly riverine nitrate concentrations using routine monitored data.

Area of Science:

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

Background:

• High-frequency nitrate (NO3-N) data are crucial for watershed management and ecosystem health.
• Accurate and cost-effective methods for acquiring high-frequency NO3-N data are needed.
• Existing methods may lack the precision or economic viability for continuous monitoring.

Purpose of the Study:

• To predict hourly riverine nitrate (NO3-N) concentrations using a hybrid deep learning model.
• To evaluate the effectiveness of routine monitored data for high-frequency NO3-N forecasting.
• To identify optimal input data strategies for improved model accuracy and stability.

Main Methods:

• Developed a hybrid Convolutional Neural Networks-Long Short-Term Memory (CNN-LSTM) deep learning model.
• Utilized routine monitored data (river depth, temperature, pH, etc.) for model training.
• Tested various input data lengths (1-30 days) and analyzed key controlling factors.

Main Results:

• The CNN-LSTM model achieved high predictive performance (Nash-Sutcliffe Efficiency 0.60-0.83).
• Models using shorter input data lengths exhibited greater accuracy and stability.
• River water level, water temperature, and pH were identified as primary factors influencing forecasting.

Conclusions:

• Deep learning, specifically the CNN-LSTM architecture, offers a powerful approach for high-frequency riverine NO3-N forecasting.
• Routine water quality data can be effectively leveraged for accurate nitrate prediction.
• Careful selection of input variables and data length is essential for optimizing model performance.