Multimodal analysis of the composition of waste-derived biomass for waste-to-energy assessment: Explainable AI (XAI) and machine learning approach

This study presents a multimodal framework based on machine learning (ML) and explainable AI (XAI) techniques for the compositional analysis of waste biomass towards an efficient waste-to-energy (WTE) system. The framework encompasses principal component analysis (PCA) for explaining the variance within the biomass-dataset, SHapley Additive exPlanations (SHAP) for interpretable ML prediction-outcomes, k -means cluster-analysis of the waste-biomass compositional profile to reveal distinct-group within the dataset, k -nearest neighbour ( k -NN) for energy-based classification, and different ML models for predicting the higher heating value (HHV). SHAP identified Carbon (C) and Hydrogen (H) as the most influential positive-predictors of HHV, while moisture-content (MC) and ash had the strongest negative-impacts. PCA captured about 78% of the variance of the dataset in 3 PCs, while k -means clustering revealed 3 distinct biomass-groups, with cluster-3 mostly suitable for direct-combustion. k -NN classifier achieved peak performance at k = 3 for classifying biomass by energy suitability, with 95.1% training accuracy. Particle swarm optimization (PSO)-optimized Artificial neural network (ANN) predicted HHV more accurately than all other models, with RMSE, MAE, and MAPE values of 3.8394, 1.5669, and 7.8502, during testing. This framework provides a scalable and interpretable tool which enables data-driven WTE planning, supports more efficient feedstock selection, optimize thermal conversion.