Improved ANN-Based Approach Using Relative Impact for the Prediction of Thermal Coal Elemental Composition Using...

Area of Science:

• Coal science and combustion engineering
• Artificial intelligence in materials science
• Chemical engineering and process optimization

Background:

• Coal properties are crucial for power plant design, efficiency, and emissions control.
• Elemental composition analysis is vital for understanding coal combustion and predicting air quality impacts.
• Existing methods for predicting elemental composition from proximate analysis have limitations.

Purpose of the Study:

• To predict the elemental compositions of thermal coals using artificial neural networks (ANNs).
• To utilize proximate analysis values as input parameters for the ANN model.
• To optimize the ANN model for improved prediction accuracy and compare its performance with existing research.

Main Methods:

• Developed and optimized an artificial neural network (ANN) model using proximate analysis data from 104 South Korean thermal coals.
• Evaluated multiple ANN topologies, activation functions (including Levenberg-Marquardt), and hidden layers.
• Assessed model performance using R-squared, Mean Squared Error (MSE), Mean Absolute Error (MAE), and compared with previous studies and adaptive neuro-fuzzy inference systems.

Main Results:

• The best ANN topology utilized the Levenberg-Marquardt activation function and 10 hidden layers, achieving the highest R-squared and lowest MSE.
• Prediction accuracy for carbon, hydrogen, and oxygen compositions was improved by 4.71-0.91% through coal rank division topology optimization.
• The ANN model demonstrated MAE accuracy improvements of 5.40% and 7.39% over previous research models and showed comparable performance to adaptive neuro-fuzzy inference systems.

Conclusions:

• The developed ANN model offers an improved and efficient approach for predicting thermal coal elemental composition based on proximate analysis.
• The study provides insights into the relative impact of ANN hidden layers on prediction accuracy for specific elements.
• Findings support the applicability and accessibility of ANNs for coal analysis in power generation, with suggestions for future research including qualitative analysis via Fourier transform infrared spectroscopy.