Intelligent Digital Twin for Optimized Oil-Water Separation Operations

Thesis Abstract

The pervasive challenge of oil-water separation in upstream and downstream operations, especially under high-viscosity, ultra-low-net-fluid production conditions, imposes substantial energy penalties and environmental liabilities; current separation facilities rely on static process configurations that fail to adapt to dynamic feed compositions and operational disturbances. This study addresses the problem by developing an Intelligent Digital Twin (IDT) framework to dynamically model, predict, and optimize oil-water separation operations, thereby reducing energy consumption, improving separation efficiency, and enabling real-time decision support for process control and asset integrity. The aim is to design, implement, and validate an IDT that integrates physics-based models with data-driven surrogates, real-time sensor streams, and advanced optimization to achieve robust performance across varying feedstocks and operating scenarios. Specific objectives include (i) constructing a modular digital twin architecture that encapsulates multiphase flow physics, separation kinetics, and chemical demulsification processes; (ii) developing data fusion and state-estimation algorithms to assimilate real-time telemetry from 60–80 sensors, including temperature, pressure, differential pressure, and emulsion quality indicators; (iii) implementing machine learning surrogates (neural networks and gradient-boosted trees) to accelerate dynamic simulations by two orders of magnitude; (iv) formulating a hybrid optimization framework combining model predictive control (MPC) with Bayesian optimization to minimize energy use and minimize emulsion carryover under uncertainty; (v) validating the IDT against a pilot-scale oil-water separator with variable feed compositions, and conducting sensitivity analyses to quantify robustness; and (vi) deriving practical guidelines for deployment in onshore and offshore facilities. The methodology adopts a mixed-methods, engineering-optimization research design. The population comprises industrial separation units and pilot-scale facilities located within a major petroleum region. A purposive sample of two pilot plants and one refinery-side separator, each instrumented with 60–80 sensors, will be used to collect data over a 12-month period. Data collection instruments include high-frequency process historians, online demulsifier dosage records, oil-water interface sensors, vibrational and acoustic emission sensors for fouling detection, and CCTV-based emulsion stability monitoring. Instrument validity and reliability will be ensured through calibration routines, cross-validation with reference lab measurements, and redundancy checks. Data analysis will employ a tiered approach (i) physics-based modeling using multiphase flow and separation kinetics; (ii) data-driven surrogate modeling via deep neural networks for transient dynamics and gradient boosting for regime classification; (iii) state estimation using extended Kalman filters and particle filters to fuse heterogeneous data streams; and (iv) optimization using a two-layer framework MPC for real-time control and Bayesian optimization for offline policy tuning. Model validation will include hold-out testing, k-fold cross-validation for surrogates, and back-testing against historical plant disturbances. Key expected findings include (i) improved separation efficiency by 6–12 percentage points under variable feed compositions, (ii) energy savings of 8–15% through optimized demulsifier dosing and temperature control, (iii) a robust fault-detection capability for fouling and sensor degradation with a false alarm rate under 2%, and (iv) demonstrable generalizability of the IDT across different feed ratios and flow regimes. The study anticipates that the hybrid IDT will outperform conventional PI and MPC schemes in both energy efficiency and resilience to disturbances, with the Bayesian component quantifying uncertainty and enabling risk-informed decisions. The study contributes to knowledge by (i) integrating physics-based models with data-driven surrogates in an end-to-end digital twin tailored for oil-water separation; (ii) proposing a modular, scalable twin architecture compatible with offshore and onshore assets and compatible with existing SCADA and OPC-UA infrastructure; (iii) advancing state-estimation and optimization under uncertainty for multiphase separation processes; and (iv) providing a transferable methodology for deploying IDTs in other petroleum processing units. The main conclusion is that an Intelligent Digital Twin can deliver real-time optimization and proactive control of oil-water separation operations under dynamic disturbances, with substantial energy and operational efficiency gains. Recommendations include expanding pilot implementations to offshore platforms, integrating real-time economics for dynamic pricing of energy and chemicals, and developing standardized interfaces for cross-plant interoperability to accelerate industry adoption.