Integrating machine learning and life cycle sustainability assessment for systematic optimization of petroleum coke oxidation for hydrogen residues processing

Abstract

This study addresses the multiphase reaction complexity and operational dynamics in fluidized bed ash-slag treatment processes by proposing a hybrid optimization framework integrating deep learning and mechanism models. A high-fidelity mechanism model, incorporating gas-solid flow, heat transfer, and reaction kinetics, was developed in Aspen Plus, generating a simulation dataset via design of experiments. A physics-constrained deep residual shrinkage network (DRSNet) was constructed by embedding mass/energy conservation equations as regularization terms, achieving precise mapping (R²>0.98) from process parameters (bed temperature, reaction pressure, fluidization air flowrate) to performance indicators (steam production, carbon content in ash-slag, carbon conversion). A multi-objective optimization model balancing economic cost, carbon emissions, and energy efficiency was solved using NSGA-II with elite strategy, yielding optimal parameters: RT=720°C, RP=3.96 bar, AF=208 t/h. Life cycle assessment (LCA) demonstrated reductions of 0.02 tCO₂eq/t steam in greenhouse gas emissions, 243 MJ/t steam in non-renewable energy consumption, and a 15.01 t/h increase in steam production compared to conventional methods. While maintaining 95% carbon conversion efficiency, the optimized process reduced non-renewable energy consumption by 14.76% and carbon emissions by 13.33%. The framework significantly improves high-dimensional optimization efficiency over traditional response surface methods while retaining accuracy. This "mechanism modeling-data-driven-intelligent optimization" paradigm offers a migratable solution for addressing "curse of dimensionality" and "model mismatch" in complex industrial processes.