A Dynamic Recurrent Neural Network for Predicting Higher Heating Value of Biomass

Area of Science:

• Biomass energy research
• Computational chemistry
• Machine learning applications

Background:

• Higher heating value (HHV) is crucial for assessing biomass energy potential.
• Existing linear models for HHV prediction have limitations due to nonlinear relationships with biomass composition.
• Nonlinear models offer a promising alternative for more accurate HHV estimation.

Purpose of the Study:

• To develop and validate a nonlinear model for predicting biomass HHV.
• To utilize Elman recurrent neural networks (ENN) for enhanced HHV prediction accuracy.
• To explore the dependency of HHV on various biomass compositional parameters.

Main Methods:

• Employing an Elman recurrent neural network (ENN) model.
• Utilizing proximate and ultimate analyses as input features for the ENN.
• Optimizing ENN architecture (hidden neurons) and training algorithm (Levenberg-Marquardt) for peak performance.

Main Results:

• An optimized single hidden layer ENN with four nodes demonstrated superior prediction accuracy.
• The ENN model accurately estimated 532 experimental HHVs with low mean absolute error (0.67) and mean square error (0.96).
• The model successfully elucidated the complex relationships between HHV and biomass components like carbon, hydrogen, and ash.

Conclusions:

• The Elman neural network provides a robust and accurate method for predicting biomass higher heating value.
• The developed ENN model enhances understanding of how biomass composition influences its energy content.
• This approach offers a valuable tool for biomass energy assessment and utilization.