A novel deep learning ensemble model based on two-stage feature selection and intelligent optimization for water...

Area of Science:

• Environmental Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate effluent total nitrogen (E-TN) prediction is crucial for wastewater treatment plant (WWTP) operational efficiency and compliance.
• The complex nonlinearity of WWTPs presents a significant challenge for multivariate time series prediction of E-TN.

Purpose of the Study:

• To develop a novel prediction framework for accurate E-TN prediction in WWTPs.
• To enhance the capture of nonlinear relationships in multivariate time series data.
• To improve feed-forward control for WWTPs, ensuring effluent compliance and reducing energy consumption.

Main Methods:

• A two-stage feature selection model was employed to identify optimal predictive variables.
• A hybrid deep learning model, CNN-LSTM-TCN (CLT), was developed by combining Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Temporal Convolutional Networks (TCN).
• The Golden Jackal Optimization (GJO) algorithm was utilized for hyperparameter optimization of the CLT model.

Main Results:

• The two-stage feature selection effectively identified an optimal subset of features for improved prediction accuracy.
• The GJO-optimized hybrid model (GJO-CLT) demonstrated superior performance across various backtracking windows and prediction steps.
• The proposed prediction system achieved excellent results in multivariate water quality time series prediction for WWTPs.

Conclusions:

• The integrated framework effectively addresses the challenges of nonlinear multivariate time series prediction in WWTPs.
• The combination of advanced feature selection and a hybrid deep learning model significantly enhances E-TN prediction accuracy.
• This approach offers a robust solution for optimizing WWTP operations and environmental monitoring.