Imperial Oil

Chemical discrimination between various formulations of aqueous film-forming foam (AFFF) is an important capability for the forensic assessment of per- and polyfluoroalkyl substances (PFAS) impacts on the environment. Identification of AFFF chemical fingerprints has been constrained by the (1) lack of detailed chemical characterizations of AFFF, (2) presence of PFAS compounds in AFFF products that are not detected and/or quantified by current analytical methods, and (3) potential chemical degradation of certain AFFF-containing PFAS towards recalcitrant end-member PFAS following release to the environment. As proof-of-concept for a technique to overcome these limitations, AFFF product samples from various manufacturers and production years were analyzed before and after application of the total oxidizable precursor (TOP) assay, which oxidizes the PFAS in a sample to their terminal end-member products. The pre- and post-TOP assay PFAS results were evaluated to determine if (1) the TOP assay generated stable and reproducible PFAS profiles, and (2) the post-TOP assay PFAS profiles enabled discrimination between legacy perfluorooctanesulfonate (PFOS) AFFF, legacy fluorotelomer (FT) AFFF, and modern FT AFFF. Principal component analysis was used to explore the post-TOP assay PFAS data to identify which variables had the strongest statistical power for AFFF product discrimination. Based on this evaluation, specific chemical ratios of the resulting post-TOP assay PFAS show potential for differentiation among legacy PFOS AFFF, legacy FT AFFF, and modern FT AFFF.

This paper introduces an adaptive weighted hybrid framework that uses a Weight Assignment Network built on the Temporal Fusion Transformer architecture. Unlike traditional hybrid models, this framework learns context-specific weighting patterns directly from model predictions. It also includes a bias correction mechanism to address systematic errors that simple weighted averaging cannot remove. We applied the framework to Primary Separation Vessels and validated its performance. The hybrid model achieved MAPE values of 0.25-3.83% in steady-state and 0.04-0.36% in dynamic mode, representing order-of-magnitude improvements over both individual data-driven and first-principle models. The attention mechanism enhances interpretability by capturing temporal dependencies and reduces input dimensionality by approximately 40% of features without compromising performance. This architecture provides a foundation for industrial deployment and maintains accuracy across different ore grades and changing conditions.

Industrial practitioners who develop linear model predictive control (MPC) applications want to prevent undesirable controller behaviour caused by ill-conditioned gain matrices and model mismatch. In this work, we propose improvements to an existing orthogonalization-based method for gain conditioning. In this offline algorithm, manipulated variables (MVs) are ranked based on their influences on the controlled variables (CVs), so that problematic MVs with correlated effects can be identified. A constrained linear least-squares optimization problem is then solved to adjust columns in the gain matrix that correspond to problematic MVs. Our goal is to update this optimization problem to prevent the optimizer from switching the signs of some gains. The updated algorithm also permits control practitioners to hold key gains constant if their estimated values are trusted. Finally, we extend the methodology to condition gain submatrices, which arise when CVs are removed from the MPC problem. An industrial fluidized catalytic cracking case study is used to test the proposed method. The conditioned gains lead to improved controller performance and less aggressive movement of MVs when there is a plant-model mismatch.

Real-time optimization of sucker rod pump operation is essential for maximizing production and reducing operational costs in oilfields. Frequent manual speed adjustments are impractical for fields with thousands of producing wells. This paper addresses these challenges by proposing a semi-closed loop system for intelligent automation of Variable Frequency Drive (VFD) operated sucker rod pumps using machine learning based fillbase prediction. The main goal is to enhance operational efficiency while mitigating risks, by incorporating human oversight for improved safety and reliability in remote oilfield operations. The proposed system integrates VFD-based speed control with ML-driven fillbase prediction (XGBoost), using dynamometer card analysis to determine pump fillage and further adjusts pump speed automatically to achieve the target fillage. A recommendation workflow monitors card quality, speed fluctuations, and alerts operators to anomalies such as persistent bad/no pump cards or significant speed changes. This ensures timely human intervention for safe operations. Various mechanisms like operator-set speed thresholds, snoozing of false recommendations, and feedback capture lead to continuous system refinement. A field trial validates this machine learning based approach and demonstrates its effectiveness in real-world conditions while maintaining operational safety and efficiency. The field trial conducted on 264 wells demonstrated significant improvements across multiple operational metrics. Production analysis revealed that 78 wells (approximately 30% of the test group) achieved an average 35% production uplift through automated speed increases, while 165 wells (62% of the test group) realized 30% power savings from optimized speed reductions. The system's ability to maintain operations around the ideal fillage or "sweet spot" not only enhanced production per unit power consumption, but also improved rod load management, thus reducing mechanical stress and extending equipment lifespan. From an operational standpoint, the solution achieved 90% reduction in manual surveillance requirements, effectively saving one full-time equivalent (FTE) workload while maintaining operational reliability. Notably, the human-in-the-loop feedback mechanism proved particularly useful, with operators intervening in only 12% of the cases where the system flagged potential anomalies or speed constrained operations. These interventions prevented potential issues while allowing the automated system to handle routine optimization. The trial also revealed that the wells operating at speed limits for extended period benefited from targeted recommendations, enabling operators to make informed adjustments. Overall, the results validate that this semi-automated approach successfully balances the benefits of AI-driven optimization with the critical need for operator oversight in oilfield operations. This hybrid approach enhances production, reduces power consumption, and mitigates operational risks, thus enabling intelligent oilfield operations. By balancing automation with human validation, the system ensures robust operations and high operational efficiency, thus marking a significant advancement in AI-augmented operational excellence in the area of oilfield automation.