Water quality prediction based on sparse dataset using enhanced machine learning.

Area of Science:

• Environmental Science
• Water Resource Management
• Machine Learning Applications

Background:

• Surface water quality is a global challenge, particularly in areas with limited monitoring data.
• Managing source-oriented pollution requires effective strategies even with infrequent data collection (e.g., weekly or monthly).

Purpose of the Study:

• To develop and evaluate machine learning models for predicting water quality using sparse datasets.
• To assess the efficacy of traditional Recurrent Neural Network and Long Short-Term Memory (LSTM) models, enhanced with the Load Estimator (LOADEST).

Main Methods:

• Investigated four machine learning models: a traditional Recurrent Neural Network and three LSTM variants, including Self-Attentive LSTM (SA-LSTM).
• Integrated models with the Load Estimator (LOADEST) for enhanced prediction capabilities.
• Evaluated model performance at a river-lake confluence with complex hydrological patterns.

Main Results:

• The SA-LSTM model, augmented with LOADEST (SA-LSTM-LOADEST), demonstrated superior performance in predicting water quality parameters like Chemical Oxygen Demand (COD_Mn) and Ammonia Nitrogen (NH₃N).
• Achieved Nash-Sutcliffe Efficiency (NSE) scores of 0.71 for COD_Mn and 0.57 for NH₃N.
• The SA-LSTM-LOADEST model reduced Root Mean Square Error (RMSE) by 24.6% for COD_Mn and 21.3% for NH₃N compared to the standalone SA-LSTM.
• Maintained predictive accuracy with extended data collection intervals (monthly vs. weekly) and forecasted pollution loads up to 10 days ahead.

Conclusions:

• The SA-LSTM-LOADEST model offers a robust approach for water quality prediction in regions with limited monitoring data.
• This method holds significant promise for improving water quality management and pollution control strategies globally.
• The model's ability to forecast pollution provides valuable lead time for mitigation efforts.