In oil refineries, production plans do not always match the actual performance of operational units, especially when feedstock quality, throughput rates, or equipment conditions change. According to Hussein Zein, vice president of Emerson in Saudi Arabia and Bahrain, this discrepancy can affect yields, product quality, and energy consumption, necessitating manual adjustments during execution.

Industrial AI models aim to reduce this variance by combining accumulated engineering knowledge with actual operational data. This approach does not replace the traditional models that refineries have used for decades, but rather enhances them to represent the nonlinear behavior of units across different feedstocks and operating conditions.

Speaking in a special interview with Asharq Al-Awsat, Zein said that traditional models often rely on linear or quasi-linear representations of unit behavior, forcing planners to work with simplified assumptions that may lose accuracy when market conditions, feedstock properties, or throughput levels change. Hybrid models, on the other hand, give operators greater ability to test multiple scenarios, adapt production decisions to variables, and reduce the time engineers spend on manual model tuning.

Hussein Zein, vice president of Emerson in Saudi Arabia and Bahrain (Image: Company)

Limitations of Traditional Planning

Refineries use planning models to convert demand forecasts, feedstock availability, operational constraints, and product specifications into executable production plans. However, the ease of using linear models in optimization does not necessarily mean they can represent all the complex interactions within units. Reactors, separators, and blending operations do not always respond linearly to changes in feedstock quality, flow rates, or catalyst condition. When a plan moves from the model to the field, operators may find that yields, product specs, or operating constraints differ from expectations.

Zein explains that this gap emerges when planning models cannot represent the true behavior of units under actual conditions. This leads to additional engineering time spent adjusting the plan during execution, or to accepting operation that is less efficient than the target level. The effects of a deviation do not stop at one unit, because the results of each unit influence feedstock blending decisions, production targets, margin forecasts, energy consumption, and coordination among interconnected units within the refinery.

Accuracy Reaching 98.5%

Emerson and Aramco reported that hybrid models achieved prediction accuracy of 98.5% in selected refining units. The figure refers to how closely the model's predicted results matched actual operational performance across a range of feedstock types and operating conditions.

Zein stresses that this percentage represents the highest level achieved in specific units—continuous catalyst regeneration units and catalytic reforming units known as Platformers—and does not represent an average for all units or operating sites. Teams are currently working to expand the same approach to hydrocracking units and test its performance there.

The importance of accuracy lies in the fact that small deviations between prediction and reality can affect yield, energy consumption, and the accuracy of production plans. Their effects also extend to feedstock blending decisions, setting production targets, and margin estimates.

Zein says that achieving prediction accuracy of 98.5% in some catalytic reforming units reduces the zone where margins are typically lost—the gap between planned performance and what is actually achieved. When the plan reflects plant behavior more closely, operations teams can commit to blending strategies and feedstock selection with greater confidence, and take advantage of optimization opportunities without costly adjustments during execution.

The data provided does not include a specific figure for the increase in yields or margins, but it links higher prediction accuracy to reducing the difference between plan and execution and improving the ability to make operational decisions.

The models achieved prediction accuracy of 98.5% in specific catalytic reforming units, and this percentage does not represent an average for all refinery units (Image: Adobe Stock)

Data-Driven Engineering

Hybrid models begin with a first-principles representation, including reaction kinetics, thermodynamics, and mass and energy balances. Then, AI uses operational data to calibrate this foundation, so the model reflects the actual unit's behavior rather than just the theoretical process behavior.

This approach combines the strengths of physical models and data-driven models. A purely physical model may be difficult to tune when trying to represent all the changing details in a real environment, while a model based solely on data may produce recommendations that violate physical laws or fail when encountering conditions not seen in the training data.

Zein believes the hybrid approach achieves "rigor and practical accuracy in one model," because engineering knowledge defines the physical boundaries, while data helps adapt the model to the unit's characteristics and operating condition.

This feature is important in complex facilities that include equipment from different generations and systems that were not originally built within a unified data architecture. It also allows refineries to use the data they have, while leveraging engineering principles to compensate for some gaps that data-only models find difficult to handle.

Human in the Decision Loop

High model accuracy does not mean transferring all decisions to automated systems. Zein distinguishes between decisions that require searching a large number of possibilities and performing repeated calculations, and decisions related to safety or exceptional circumstances. Optimizing feedstock blending, multi-period production planning, and routine model maintenance are areas where AI can support or automate parts, while operators and engineers continue to supervise.

In contrast, decisions related to personnel safety, environmental risks, or unusual operating conditions should remain under direct human control. Zein says the required balance is that "AI should expand what engineers can achieve, without transferring responsibility for critical judgments away from the people who understand the plant."

When a system recommendation conflicts with an expert engineer's judgment, the situation should not be treated as a conflict between human and machine. The discrepancy may reveal a factor the model did not include, or information the engineer possesses that has not yet been translated into rules or data.

The vice president of Emerson in Saudi Arabia and Bahrain states that "the conflict between an AI recommendation and engineering judgment is a valuable signal, not a problem to be solved by choosing one side." If the engineer's reservation reveals a real shortcoming, the model should be updated. If the system shows an opportunity that was not obvious, the discrepancy becomes a learning and review process.

Human oversight remains essential for decisions related to safety, environment, and unusual operating conditions (Image: Adobe Stock)

Accuracy Changes Over Time