Interpretable machine learning-assisted advanced exergy optimization for carbon-neutral olefins production

Abstract

The CO₂-to-light olefins technology represents a significant approach to mitigating the greenhouse effect and advancing green energy solutions. However, little literature comprehensively analyzes and optimizes its thermodynamic performance. This study proposes an interpretable machine learning-assisted advanced exergy analysis and optimization framework to ascertain the actual improvement potential and determine effective strategies for optimizing this system. The advanced exergy analysis method aims to identify the avoidable exergy destruction and interactions between components of the system, while integrating an interpretable machine learning model to provide the key parameters for enhancing the system's exergy efficiency through feature importance analysis. The findings indicate that the exergy destruction of the system amounts to 656.06 MW, with 96.81 % of this exergy destruction being attributed to endogenous factors and approximately 66.51 % of it being potentially avoidable. The random forest model, exhibiting superior predictive accuracy compared to other machine learning models, is coupled with the interpretable Shapley additive explanation approach to discern the most crucial parameters of the system. Results indicated catalyst properties have the greatest impact on the output performance of the system, contributing up to 66.1 % to the predicted results. The active component type, reaction temperature, and promoter content have the largest contribution to the prediction of CO₂ conversion ratio and light olefins selectivity. Furthermore, the key input features are optimized by screening for better catalysts and conducting sensitivity analysis. After optimization, the system's avoidable exergy destruction is significantly saved by 32.27 %, resulting in an enhancement in exergy efficiency by 8.12 %.