Tao Liu
Technology Planning Coordinator
Dalian West Pacific Petrochemical Co., Ltd.
Dr. Liu has nearly two years of experience in the energy industry. Currently serving as the Technology Planning Coordinator at Dalian West Pacific Petrochemical Co., Ltd., he holds a Ph.D. in Chemical Engineering Technology from Tianjin University. His research focuses on intelligent transformation applications in the refining and chemical industry. Notably, he presented an award-winning conference paper at the China PetroChina Grand Model Conference.
Participates in
TECHNICAL PROGRAMME | Energy Technologies
Smart Infrastructure for the Future Energy Industry: Digitalisation & Innovation
Forum 18 | Digital Poster Plaza 4
27
April
15:30
17:30
UTC+3
Digital twin technology is crucial for refining enterprises aiming for operational excellence, yet traditional mechanistic and data-driven models struggle to capture the complexities of refining processes. A dual-driven digital twin that integrates data and theoretical frameworks is essential for achieving a balance of interpretability, accuracy and adaptability, as it embeds physicochemical constraints into machine learning architectures to ensure predictions adhere to conservation laws while dynamically learning unmodeled phenomena. Despite its theoretical promise, implementing a dual-driven digital twin in refining systems faces three critical challenges. One significant issue is high-quality dataset engineering due to diverse data sources causing issues like missing values and temporal misalignment, necessitating robust preprocessing techniques. Another critical challenge is achieving real-time high-fidelity modeling, which requires multi-scale approaches that integrate molecular-level kinetics and fluid dynamics while utilizing reduced-order models and neural network pruning for efficient inference. Furthermore, online optimization and decision-making are complicated by high-dimensional, non-convex objectives, necessitating a closed-loop “sense-decide-act” framework that allows reinforcement learning agents to dynamically adjust operating parameters while maintaining safety margins.
This study addresses these challenges through a novel methodology combining domain knowledge embedding and holistic optimization. Initially, high-quality datasets are developed through expert and theory-guided protocols that include outlier detection and imputation using first-law, K-means clustering combined with domain expertise to define parameter correlations, and automated cloud simulation platforms that generate physics-compliant samples. Secondly, an unified computational architecture integrates multi-scale constraints, such as molecular-scale reaction kinetics and equipment-scale hydrodynamics, into loss functions using Lagrange multipliers. Continuous online learning allows for the adaptation of model parameters to real-time sensor data, achieving less than 2% prediction errors in product properties. Thirdly, a holistic optimization algorithm based on gradient descent synchronizes operational variables, including distillation cut points and pump frequencies, thereby reducing optimization cycles from hours to minutes. Implemented in a 10-million-ton-per-year atmospheric-vacuum distillation unit, this framework facilitated closed-loop operation, resulting in a 1.8% increase in light oil yield and generating annual economic benefits exceeding CNY 25 million. This work illustrates that the fusion of theory-constrained data and cross-scale modeling is essential for the advancement of next-generation intelligent refineries.
Key words: Digital twin, data and theory, high-quality datasets, refinery optimization.
Co-author/s:
JiaHua Zhang, Dalian West Pacific Petrochemical Co., Ltd.
This study addresses these challenges through a novel methodology combining domain knowledge embedding and holistic optimization. Initially, high-quality datasets are developed through expert and theory-guided protocols that include outlier detection and imputation using first-law, K-means clustering combined with domain expertise to define parameter correlations, and automated cloud simulation platforms that generate physics-compliant samples. Secondly, an unified computational architecture integrates multi-scale constraints, such as molecular-scale reaction kinetics and equipment-scale hydrodynamics, into loss functions using Lagrange multipliers. Continuous online learning allows for the adaptation of model parameters to real-time sensor data, achieving less than 2% prediction errors in product properties. Thirdly, a holistic optimization algorithm based on gradient descent synchronizes operational variables, including distillation cut points and pump frequencies, thereby reducing optimization cycles from hours to minutes. Implemented in a 10-million-ton-per-year atmospheric-vacuum distillation unit, this framework facilitated closed-loop operation, resulting in a 1.8% increase in light oil yield and generating annual economic benefits exceeding CNY 25 million. This work illustrates that the fusion of theory-constrained data and cross-scale modeling is essential for the advancement of next-generation intelligent refineries.
Key words: Digital twin, data and theory, high-quality datasets, refinery optimization.
Co-author/s:
JiaHua Zhang, Dalian West Pacific Petrochemical Co., Ltd.
