AI-driven Digital Twin Framework for Predictive Maintenance, Asset Integrity Management, and Energy Optimization in Smart Oil and Gas Industrial Systems

Abstract

Advances in AI, physics-informed modelling, and Digital Twin technologies offer the potential to enhance predictive maintenance, asset-integrity assessment, and energy management within oil and gas processing systems. However, these frameworks are most useful when they are based on a common data source, use a consistent model configuration, and are tested under a defined set of operating conditions. This study presents and evaluates an AI-assisted, physics-informed DT for an integrated crude distillation unit and hydrotreating system based on a 180-day synthetic operating dataset. The framework integrates a bidirectional Long Short-Term Memory (BiLSTM) network for equipment-anomaly detection and remaining useful life (RUL) estimation, a Physics-Informed Neural Network (PINN) to predict the wall thickness of process piping, and the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to optimise energy. During the independent test period, the precision and recall of the BiLSTM were 0.9654 and 0.9941, respectively, with a false-alarm rate of 3.85% and a RUL mean absolute error of 6.24 days. The RMSE and MAE for the independent evaluation period of the wall-thickness were 0.00585 mm and 0.00526 mm, respectively, for the PINN. It was found that a feasible operating condition could be obtained from NSGA-II using a load setpoint of 0.88 and a recycle fraction of 0.02. The Digital Twin strategy reduced simulated electricity consumption by 2.24% and fuel-energy consumption by 2.14%, equivalent to an estimated reduction of 452.67 t CO₂e over the 180-day simulation period when compared to the defined baseline. The results show that predictive maintenance, physics-informed asset-integrity prediction and energy optimisation can be integrated computationally within a single Digital Twin framework in a viable manner. The results do not constitute field validation, as the data are synthetic and the energy and environmental estimates are model-based. Further assessment using plant-specific sensor, maintenance, inspection, and energy-consumption data is needed before industrial implementation.