• What do you see necessary for successful implementation of plant-wide AI and machine learning (ML) in the refinery and petrochemical complex? Can AI/ML strategies provide additional benefits beyond APC?



  • Nagarathinam S Murthy, McKinsey & Company, Chennai, nssvdvr@gmail.com

    The simple answer to this tough question, in my view, will be look into the variability in output indicators like conversion, quality consistency, deviation from assigned IOW (integrity operating window) and BOZ (best operating zone), unexpected catalyst deactivation, excessive fouling and or pressure drop increase, etc. Use statistical test by taking any one of the output variables mentioned above using CpK value, precision and tolerance allowable. One can decide the need for advanced mathematics like AI / ML.



  • Abishek Mukund, Imubit Inc., abi.mukund@imubit.com

    There are 2 requirements that are needed to leverage AI/ML plant-wide for bottom-line value improvement: full alignment upfront on the operational strategies and the ability for the entire organization to understand the AI/ML solution.

    It takes all the plant perspectives to truly understand the scope of the optimisation objective. This means aligning planning economics, process control, process engineering, and operations on the operating strategy and economic benefits. Once the plant groups are aligned on the scope, AI implementation can follow a streamlined, resource-efficient process.

    Any new technology needs full organizational adoption to create sustained value. Often, different solutions are only understood by a single plant group. AI has an even greater challenge of explaining what models have learned from data rather than pure first principles. Today, there is a separate field of AI called Explainibility where leaders in AI are discovering ways to convey what an AI model has learned for those who only have domain knowledge. AI solutions must offer intuitive ways to explain what the models have learned to allow plant operators, engineers, and leaders to build confidence in data-driven models. AI Explainibility solutions are a must if organizations want to adopt AI and sustain benefits.
    APC has added significant value for stable control where linear assumptions are close to reality. Decades of continued refinement has incrementally improved APC value add, but specific forms of AI add a step change in plant optimisation and control. Deep reinforcement learning models have been able to work with and without APC to directly control the plant to improve margins while respecting safe operating limits and sustainability objectives. Use of deep learning neural networks allow engineers to model the plant’s true relationships, extracting non-linearity and time dynamics from the years of historian data. Then, use of deep reinforcement learning (DRL) allows models to develop human-like intuition to continuously optimise and control in the dynamic plant environment.

    We have had tremendous success implementing DRL models on top of APC where APC manages lower-level constraints but needs direction to manage larger control and optimisation objectives. This includes non-linear relationships in conversion units and complex objectives across a system of units with varying time dynamics.

    With the requirements met and the right AI for the desired objective, AI can capture previously unrealized value.



  • Doug Morgan, Searles Valley Minerals, morgan@svminerals.com

    I concur with Ron Beck. AI and Evolutionary (Genetic) Algorithms applied to planning is where the easy wins are.  My experience with AI in APC was extremely valuable - in coming to a better understanding of my role in capitalising on the intelligence, adaptability, and resiliency of our human operators.  

    The example I like to provide is NOx management on a coal-fired power plant (something found in many refineries back in the day). Given the complexity and lack of first-principle understanding a collection of something like 21 variables in the form of various dampers, fans, pivoting nozzles, and fuel feed modulators imparts on NOx output, one would think AI could make an operator obsolete. However, try this... Go to your power plant and ask the operators to change NOx output by 10% one way or another without bumping steam or power production. Then ask yourself if you could justify the work in getting a computer to do what they just did - perfectly.  Now realize that all that intelligence isn't worth much if your instruments and/or analyzers are not providing the right data or their supervisor is interfering with counter production parameter dictates.



  • Alvin Chen, BASF Corporation, alvin.chen@basf.com

    Successful implementation of machines with artificial intelligence (ML/AI) in refinery or petrochemical complexes requires a deliberate and thoughtful targeted approach with clearly defined benefits, a robust and safe technology framework, and a clear economic benefit. 

    These hydrocarbon processing facilities have some of the highest safety standards in the world, and the use of their technologies in ML/AI offers the potential for benefits, yet they must achieve standards of safety and robustness that are often higher than many other industries. This can often set the speed of adoption in the industry, yet it cannot be missed as a tool because there are large potential benefits to operations that can be achieved with AI/ML.

    ML/AI offers a chance to continuously learn and improve, leading to better productivity, less downtime, and ultimately improved cost efficiency. Specific AI/ML strategies can provide additional benefits beyond automated process control (APC) by extending analysis to monitoring market conditions and adjusting crude acquisition plans.
    Additionally, it can allow catalyst formulation adjustments to accommodate market changes and uncertainty. A specific example is changing catalyst to move away from C₄= selectivity in favour of C₃/C₄ flexibility when there is uncertainty about which production will have the highest value and demand.

    An AI/ ML program can help in these types of decision-making recommendations while allowing a final human interface in the process. This is just one simple example of how these tools can be implemented, which is why today’s much more complex logistics, process control, and maintenance monitoring are being improved through AI/ ML.



  • Damien Maintenant, Axens, damien.maintenant@axens.net

    By unifying planning/scheduling and APC and coordinating the controllers’ objectives, plantwide optimisation provides additional benefits beyond APC. AI/ML techniques can be used to simplify plantwide optimisation tools that are complex to operate and strenuous to maintain, but these algorithms need a large amount of data and have to be developed, guided, and monitored by engineers with rigorous knowledge of the process and the operation.

    In addition, in an increasingly connected global market context, refining and petrochemical schemes are more and more complex and integrated, complicating plantwide optimisation. As a result, this is not only a matter of data science but also of process expertise, as it is of real importance to understand the interactions between the different processing units across the plant. Regarding tools and techniques, hybrid modelling, meaning a combination of historical operating data and first-principles models, must be taken into consideration for developing such solutions.

    It is worth mentioning that some mandatory project phases must be respected. The first milestone is to express the objectives of the optimisation solution clearly; the second milestone is to accordingly define the strategy to achieve these objectives and make available all needed resources. Consider that agility is also key to redefining objectives or resizing resources as necessary. Agility is also a way to extend the scope of the solution over time by starting plantwide optimisation, for instance, of the utilities and hydrogen network, then including  pools management, and so on.

    In conclusion, data availability and monitoring, professional expertise (process, operation, control, data science), efficient project management, and resources (on the one hand, people and, on the other hand, the tools and techniques) are necessary for the successful implementation of plantwide optimisation using AI/ML.



  • Ron Beck, AspenTech, ron.beck@aspentech.com

    AI and machine learning are positioned to create very significant benefits in operational excellence. In fact, some of the benefit areas will be significantly greater than those that may come from APC.

    One area we have already implemented is in applying AI/ML to refinery and olefins planning. AI can augment the existing planning systems by helping planners sift through hundreds of complex plan alternatives to find those most likely to achieve lower carbon operations based on results from historical planning periods.

    Another area is achieving more accurate digital twin models by applying AI to data streams to improve results from rigorous models. These will help in crucial areas of energy efficiency, such as heat exchanger and process unit fouling. By being able to model the actual conditions, not the ideal ones, these AI-based hybrid models will have a big impact on optimising operations for energy efficiency.

    It is necessary to have technology that democratises the application of Industrial AI, making it possible for AI and ML to be used by engineers of different skill levels.
    Very often, process data is treated as company intellectual property. As a result, access can be on a need-to-know basis. There must be documented processes for people who need that access to be able to gain it. Likewise, it is necessary to have good data quality. There may be time periods requiring users to ‘condition’ data to be able to get the most out of it, for example, where:
    - The process is not steady (assuming you are calibrating a steady-state model)
    - The process measurement is bad (for example, outside the measurable range of the instrument)
    - The data does not close the material balance
    - The plant is operating in a regime that is out of the norm.

    There are many benefits to using AI/ML. Aspen hybrid models combining first principles and AI/ML allow engineers to:
    - Capture unknown or unmeasurable details of phenomena while recalibrating models to changing process conditions more easily with AI/ML
    - Represent real plant behaviour with models created from operational data and first-principles constraints
    - Create high fidelity models that can be used for rapid and accurate decisions in engineering and operations or to expand modelling scope
    - Better planning decisions to recapture benefits
    - Easily incorporate complex process units into the scope of closed-loop optimisation.