Refinery Advisor Solution Optimizes the Commercial Plan Using Cognitive Hybrid Intelligence

Summary

For process manufacturers, solving the optimization challenge and improving margins is paramount. In the downstream energy industry, model-based solutions have proliferated. Yet their use and value are impeded by manual processes, and inefficient and untimely coordination between schedulers, process engineers, and operators. Significant opportunity exists to close the optimization gap between the refinery plan-versus-actual by codifying expert knowledge and automating its broader use via artificial intelligence.

Model-based technologies have a long history and continued success addressing process variability, constraints and optimizing for a specific objective function. The emergence of artificial intelligence in industrial manufacturing has become a “catch-all” for a variety of methods and approaches, such as machine learning, within the spectrum of AI, confusing asset owners about how to apply the correct AI methods to improve well-known processes.

Beyond Limits LLC recently briefed ARC Advisory Group on their LUMINAI Refinery Advisor application. The application is built using cognitive hybrid intelligence that helps to address the optimization gap by focusing on problems that require interpretation and complex reasoning between the scheduler, process engineer, and operator. Our key findings of the briefing include:

• Process interruptions, delayed startups, process variability, or other unexpected problems, often create a gap between planned vs. actual.
• Cognitive hybrid intelligence combines numeric AI with symbolic AI that solve general classes of problems using human-like reasoning to produce actionable intelligence.

• Beyond Limits LLC Refinery Advisor is built on knowledge and expertise that provides real-time guidance to close the plan-versus-actual gap using an iterative approach.

Traditional Refinery Optimization Practice

Today, there is a tremendous dichotomy between optimization using real-time control and automation and manually gathering and communicating information to operators. While advanced process control (APC) and dynamic optimization are “table stakes” for real-time control, many refinery planners, schedulers, and optimization engineers still work independently using different data models with very little data integration. Even with these isolated practices, they have been reasonably successful at helping operators keep units running to target. However, the opportunities to improve margins are significant. Planners will often run an off-line linear program (LP) optimization model to develop an average operating plan across the planning period (i.e., weeks and months into the future). Schedulers will break down the current planning period into a sequence of scheduled steps and work with a process engineer to ensure feasible management of inventories.

This approach leaves operations with a plan based on average process behavior, but one that’s less than optimal. The daily schedule is feasible, but not necessarily optimized each day. The manual exchange of information between process engineer, scheduler, and operator leaves room for error. In many cases, the schedule cannot be achieved due to process variability, mechanical issues, or unplanned asset failures or when portions of the refinery are constrained. Often, the available corrective actions may not always be apparent or within the comfort zone of an operator.

After accounting for process interruptions, delayed startups, process variability, operating outside of optimal conditions or the time delay with operators researching or asking advice from optimization engineers, there’s always a gap between plan vs. actual, often leaving a significant opportunity to improve profitability. This gap is mainly due to modeling inaccuracy in both planning/scheduling and control models, and mismatches between planning, scheduling, and control constraints and targets and the time operated in an off-plan state. The application of AI is proven to be successful in improving the accuracy of APC and optimization models to address real-time control and optimization. However, another AI method – cognitive hybrid intelligence - is also helping to address the optimization gap by focusing on problems that require interpretation and more complex reasoning between the scheduler, process engineer, and operator.

Enter Cognitive Hybrid Intelligence

Many forms of AI are applicable to industrial manufacturing. Within the AI spectrum, numeric AI is a class of statistical methods, trained using sample data to label specific sets of objects and conditions, and a result interpreted to produce an actionable outcome, and often used for failure prediction of an asset. Cognitive AI are analytical methods designed to replicate human thought processes in computers and machines. Cognitive hybrid intelligence combines numeric AI with knowledge-based systems that solve general classes of problems using human-like reasoning to produce actionable intelligence. This adds problem-solving capabilities to understand and manipulate labels to answer questions using “book-like” or tacit knowledge to educate the system with algorithms, rules, strategies, and workflows.

Cognitive systems require a human-like understanding of complex domains capable of adapting to uncertainty in both their knowledge and data, supporting their answers with a human-understandable audit trial. Cognitive systems require integrating different learning and adaptation techniques to overcome the limitations of the individual technologies and achieve synergetic effects through hybridization or fusion of the knowledge-based and numeric technologies. Machine learning alone is often insufficient when results must be explainable, and data is limited, unreliable, or misleading. Likewise, knowledge-based reasoning alone is inadequate when data sets are large, and data reduction is needed to make them solvable.

Cognitive systems represent knowledge and wisdom in a variety of forms, educated from domain knowledge and experience. It can reason through general classes of problems, (not limited to specific use cases) and transform numeric-only models that describe a situation into intelligent agents that can adjust and dynamically recommend action that is auditable with the source of how the outcome was reached. Unlike numeric AI, which requires a considerable amount of data to build a reasonable model, cognitive hybrid approaches require less data and potentially a shorter implementation time.

Applying Cognitive Intelligence to Refinery Operations

Beyond Limits LLC, an industrial and enterprise-grade AI technology company, recently briefed ARC Advisory Group on their LUMINAI Refinery Advisor solution. According to the company, Refinery Advisor is a no-code SaaS, cloud-based, AI-driven decision support system that maintains a symbolic representation of a campaign and refinery plan. The solution captures the expertise and knowledge of highly skilled process engineers and operators and makes timely and transparent recommendations that enable operators to close the plan-versus-actual gap and maximize operational efficiency and refinery margins.

The LUMINAI Refinery Advisor provides the operator with the correct information and instructions in the exact language that the engineer wants to communicate (which is often unique to a site or company). The activity typically involves the process engineer physically meeting the board operators, armed with the knowledge of what's going on with a process asset (distillation tower or reactor) and articulating the steps necessary to correct any off-plan process conditions. Tacit process knowledge is embedded in the Refinery Advisor application with its history because it is now machine-readable.

The advisor application leverages these identified best practices, strategies, and actions to guide off-plan identification to resolution using symbolic AI (high-level symbolic representations of problems, logic, and search). It can identify deviations from the plan and provide optimal recommendations to operators in real time. The reasoning engine can represent sensor data as symbolic states (i.e., above limit, volatile or at target) and encode process engineers' knowledge using a symbolic plan representation. Modular agents with local configurations by process engineers can parse and prioritize the execution of various conditions using numeric data from a process historian or time-series data, and represents these as symbolic states. The Advisor engine can then monitor the process states every 15 minutes using “encoded knowledge“ to identify off-plan performance, prioritize, evaluate, and reason through states to solve as a process engineer would.

Furthermore, the Advisor provides iterative feedback as to when operators took opportunities to optimize the process based upon advice from the system and when they could not because of existing constraints. This retro analysis post-campaign provides the ability for process engineers to go back and improve the process, and justify capital improvements as necessary, without gathering data from multiple sources.

According to the company, one problem with using a strictly numeric AI approach is that algorithms tend to be very hypersensitive to the training data and changing process conditions. The Refinery Advisor cognitive system eliminates that issue because it is built to take knowledge and wisdom in a variety of forms to be able to reason through various general classes of problems. Beyond Limits doesn’t do away with ML methods completely, but instead keeps ML components in the background. ML essentially characterizes the raw data into a set of features, at which point the symbolic components can operate on that data rather than having that being the primary decision-making method. That’s the cognitive hybrid intelligence approach.

Conclusion

The Beyond Limits Refinery Advisor solution may be a valuable tool for energy and chemical companies to close the gap between plan-versus-actual and complement existing model-based automation solutions. Cognitive systems are designed for imprecisely stated problems containing missing or misleading data and situations requiring human-like insight when an answer is not apparent. Machine learning has become synonymous with AI but machine learning, as with other conventional numeric AI techniques, exhibits rigidity, brittleness, cannot solve certain problems, and may lack explainability. Machine learning alone has difficulty making connections when it's not trained, since it doesn't have the capacity to reason.

As such, real-time model-based solutions, such as advanced process control (APC) and dynamic optimization real-time control and automation, remain as valuable tools in the optimization toolkit. Others have attempted to create operator advisor solutions using rules and basic engines; however, these have been error-prone and difficult or expensive to maintain.

Cognitive systems address the problems that require interpretation in a dynamic environment such as a continuous process environment,  taking the expertise from the best engineer or operator who made the correct interpretation of that problem, therefore pushing out knowledge to places and people, where it can be implemented more effectively.

While the Refinery Advisor application is built specifically to optimize the Refinery plan, ARC believes the Beyond Limits cognitive hybrid intelligence technology will be well-suited for a variety of business processes and industries.

ARC Advisory Group clients can view the complete report at  ARC Client Portal

If you would like to buy this report or obtain information about how to become a client, please  Contact Us

Keywords: Cognitive Analytics, Beyond Limits, Artificial Intelligence (AI), Advanced Analytics, Plan versus Actual, Refinery Margin Improvement, ARC Advisory Group.