SHAP-Explainable Ensemble Machine Learning for 4E Analysis and Multi-Objective Optimization of a Biomass-Fired Gas Turbine Integrated with ORC and Absorption Chiller Multigeneration Syste

Abstract

Biomass-fired multigeneration systems that utilize gas turbines coupled to organic Rankine cycles (ORC) and absorption chillers (AC) show great potential for providing clean and reliable energy. However, their complex non-linear relationships among decision variables (coupled thermodynamics and economics) and the black-box nature of most machine learning (ML) surrogates hinder these systems’ multi-objective optimization. In this paper, we develop a SHAP-explainable ensemble ML framework to enable 4E (energy, exergy, economic, and environmental) analysis and multi-objective optimization of a biomass-fired gas turbine/AC/ORC-based multigeneration system. A dataset of 1,000 Latin Hypercube Sampling (LHS) simulation cases from a validated thermodynamic model is generated and five ensemble models (RF, GBR, XGBoost, LightGBM, CatBoost) are trained, compared, and evaluated, with a tuned XGBoost achieving R² > 0.989 for all targets. Multi-level SHAP analysis identified turbine inlet temperature and pressure ratio as the most important design drivers for all performance metrics. Multi-objective optimization with NSGA-II and multi-criteria decision-making with TOPSIS identified the optimal design with an exergy efficiency of 61.8% and SUCP of ~7.8 $/GJ. The presented framework, which overcomes the black-box limitation and achieves high predictive performance, can be used to bridge the gap between predictive accuracy and engineering interpretability in the design of biomass-fueled multigeneration systems.