A novel RF-CEEMD-LSTM model for predicting water pollution.

Area of Science:

• Environmental Science
• Water Quality Management
• Data Science

Background:

• Accurate water pollution prediction is crucial for effective water environment prevention and control.
• Challenges in prediction include input variable uncertainty and the non-stationary, nonlinear nature of pollution data.
• Existing models often struggle with the complexity of water pollution time series.

Purpose of the Study:

• To develop a novel water pollution prediction model (RF-CEEMD-LSTM) that improves prediction accuracy and reliability.
• To combine the strengths of Random Forest (RF), Complementary Ensemble Empirical Mode Decomposition (CEEMD), and Long Short-Term Memory (LSTM) models.
• To analyze the driving forces influencing water pollution prediction.

Main Methods:

• Proposed a hybrid model: Random Forest (RF) combined with Complementary Ensemble Empirical Mode Decomposition (CEEMD) and Long Short-Term Memory (LSTM).
• Utilized measured data for experimental validation.
• Performed driving force analysis to identify key pollution variables.

Main Results:

• The RF-CEEMD-LSTM model achieved a Mean Absolute Percentage Error (MAPE) of less than 8% for water pollution prediction.
• Significantly reduced Root Mean Square Error (RMSE): 62.6% vs. LSTM, 39.9% vs. RF-LSTM, and 15.5% vs. CEEMD-LSTM.
• Total Nitrogen (TN) was identified as the most significant driving factor in water pollution prediction.

Conclusions:

• The proposed RF-CEEMD-LSTM model demonstrates superior performance in predicting non-linear and non-stationary water pollution sequences.
• The hybrid approach effectively addresses the complexities of water pollution data, enhancing prediction accuracy.
• Findings provide valuable insights for identifying key water pollution variables and advancing prediction methodologies.