Multivariate and Machine Learning Analysis of Biomass Combustion Regimes in a Bubbling Fluidized Bed

This work investigates the application of multivariate statistics and machine learning to analyze combustion and emission behavior of different biomasses in a bubbling fluidized bed combustor. Raw, leached and blended biomasses, including sawdust, elephant grass, palm fiber and empty fruit bunches, were evaluated using experimental process data related to temperature, pressure, feeding rate and flue gas composition. After data cleaning through boxplot and Mahalanobis distance methods, Principal Component Analysis (PCA) was applied to reduce dimensionality and identify the most relevant variables controlling combustion stability and pollutant formation. The first two principal components explained 83.1% of the total variance, highlighting the influence of bed temperature, freeboard temperature, furnace pressure and gas composition. Cluster analysis indicated seven operational regimes, validated by Silhouette, Davies–Bouldin and Calinski–Harabasz indices. Supervised classification showed that Linear Discriminant Analysis achieved the best performance, with macro-F1 of 0.97 and Cohen’s kappa of 0.96. The results demonstrate that the PCA–k-means–LDA framework is a robust and interpretable tool for identifying efficient combustion windows, classifying biomass behavior and supporting emission mitigation strategies for NO., SO2 and CO.