An Ensemble Deep Belief Network Model Based on Random Subspace for NO _x Concentration Prediction.

Area of Science:

• Environmental Science
• Computer Science
• Engineering

Background:

• Effective prediction of nitrogen oxides (NOx) emissions is crucial for reducing air pollution.
• Traditional NOx modeling methods often struggle with complex nonlinear relationships and generalization.
• Real-time prediction capabilities are essential for dynamic emission control strategies.

Purpose of the Study:

• To propose a novel real-time NOx prediction model using an ensemble deep belief network (DBN).
• To enhance the accuracy and generalization ability of NOx emission prediction models.
• To provide a robust tool for managing and reducing NOx emissions from industrial sources like boilers.

Main Methods:

• Variable importance projection (VIP) analysis for feature selection and time delay estimation.
• Phase space reconstruction of historical data for capturing system dynamics.
• Ensemble strategy using a random subspace method with Deep Belief Networks (DBNs) as submodels and a Back Propagation Neural Network (BPNN) for combination.

Main Results:

• The ensemble DBN model effectively captures the nonlinear relationships between input variables and NOx concentration.
• The proposed model significantly outperforms commonly used methods like BPNN and Support Vector Machines (SVM) in NOx prediction.
• Demonstrated superior prediction performance and generalization ability for NOx emission modeling in a 660 MW boiler.

Conclusions:

• The ensemble DBN model offers a powerful and accurate approach for real-time NOx emission prediction.
• This method provides a significant advancement over existing techniques for environmental monitoring and control.
• The findings support the application of advanced ensemble deep learning techniques for industrial pollutant emission management.