Self-Evolving Hybrid Digital Twins Using Continual Learning and Edge AI for Adaptive Refinery Optimization Under Dynamic Operational Uncertainty

The petroleum refining industry faces escalating operational complexity driven by volatile feedstock compositions, fluctuating energy costs, tightening environmental regulations, and increasing demands for product flexibility. Conventional process optimization frameworks whether first-principles models or static data-driven surrogates—suffer from model drift as operational conditions evolve, requiring costly periodic retraining cycles and human intervention. This paper introduces a novel architecture for Self-Evolving Hybrid Digital Twins (SE-HDT) that integrates continual machine learning, physics-informed neural networks (PINNs), and edge AI computing to enable autonomous, real-time adaptive optimization of refinery operations under dynamic uncertainty. The proposed framework combines a high-fidelity mechanistic core with an adaptive neural surrogate layer that continuously updates its internal representation without catastrophic forgetting, using elastic weight consolidation (EWC) and progressive neural network expansion. Edge AI inference nodes deployed at critical process units enable sub-۱۰۰ ms optimization decisions independent of cloud connectivity, while a federated learning backbone allows knowledge transfer across geographically distributed refinery sites without compromising proprietary process data. The SE-HDT framework was validated on a large-scale integrated crude distillation unit (CDU) and fluidized catalytic cracker (FCC) complex over an eighteen-month pilot deployment at a ۲۰۰,۰۰۰ barrel-per-day (BPD) refinery. Results demonstrate a ۷.۳% reduction in energy intensity (GJ/tonne), a ۴.۱% improvement in high-value product yield, a ۹۹.۹۷% model uptime despite three major feedstock transitions, and a ۶۲% reduction in manual model maintenance interventions compared to baseline static surrogate approaches. This work advances the frontier of autonomous process systems engineering and provides a deployable blueprint for next-generation intelligent refineries.

کلیدواژه ها:

digital twin ، continual learning ، edge AI ، refinery optimization ، physics-informed neural networks ، federated learning ، elastic weight consolidation ، process uncertainty ، adaptive control