Interpretable machine learning models for forecasting elemental composition of solid biomass fuels from proximate analyses in energy processes

Abstract

Biomass is a renewable and sustainable source of green energy. A key factor in evaluating its energy potential is its elemental composition—primarily carbon (C), hydrogen (H), and oxygen (O). This information is vital for accurate material balance calculations, efficient design and operation of combustion systems, and determining oxidant requirements for combustion and gasification. It also enables prediction of gas composition from these processes. While ultimate analysis provides this elemental data (i.e., carbon (C), hydrogen (H), oxygen (O), nitrogen (N), and sulphur (S)), it is expensive and time-consuming. In contrast, proximate analysis is simpler, offering data on moisture content (MC), volatile matter (VM), ash content (Ash), and fixed carbon (FC). Predicting elemental composition from proximate analysis requires robust models. This study develops nonlinear predictive models using two machine learning (ML) techniques: multilayer perceptron (MLP) and a hybrid approach combining the Harmonic Search Optimization Algorithm (HSOA) with Multivariate Adaptive Regression Splines (MARS). Based on a dataset of 203 biomass samples, six ML models were built to estimate C, H, and O content. Results show these ML models outperform traditional linear models in accuracy and generalizability. Specifically, the optimal HS/MARS models achieved coefficients of determination of 0.8339, 0.8676, and 0.8714 for C, H, and O, respectively. The HS/MARS models also outperformed the MLP models, demonstrating their superior predictive capability.