By Allison Buenemann, Product Marketing Manager, Imubit
Process and control engineers ask detailed questions about how closed-loop control works in a plant. We collected the ones that come up most often in our webinars and answered them here, grouped by architecture, applications, continuous improvement, technology comparison and people.
Architecture
How is the software structured, and how is data shared across teams?
There are two main components. The cloud-hosted Imubit Platform gives every team a shared environment to build, evaluate and track the performance of closed-loop models. The On-Premise Application runs the closed-loop Deep Learning Process Control® (DLPC®) controller at the site, and it can be deployed in the DMZ or on the process control network (PCN).
There is no two-way exchange of data between the on-premise controller and the cloud database. The On-Premise Application communicates with sources on site, including the DCS, the laboratory system and the data historians, over OPC and OPC UA, the industry-standard protocols.
Does the runtime controller run on the process control network?
The On-Premise Application running DLPC® connects either to the DCS or to the existing APC, and it sits at the PCN level to communicate with them. Data extraction from the site historian for model training is typically done from the business network.
How does the optimization get economic, pricing and cost data?
Some plants push prices from their LP to the historian or straight to the DCS. In that case the controller picks up price-deck updates as they happen. Other sites prefer to control when updates take effect, and they enter price changes manually in the On-Premise Application. Either way, the controller optimizes against the prices it has, so the more often prices are updated, the better it can optimize.
How often is data collected, and how often are setpoints updated?
This is configurable for each application. Data is typically collected every few minutes, and across applications the read and write frequency ranges from one minute to an hour.
Is there automated feedback from lab and analyzer results to keep inferentials current?
Yes. Inferentials are biased on site using a cumulative sum (CUSUM) method or a straight linear bias, so the model stays current with slow changes in unit operation and feed quality. Imubit's implementation engineers set this up with the site team. The approach is similar to how other technologies bias their inferentials to give priority to the most recent lab results.
Applications
Has Imubit solved similar FCC optimization problems at other refineries?
Yes. FCC units are among the most common DLPC® applications. The strategies span every stage of FCC operation, including feed preheat, fractionation, the reactor and regenerator, the C3/C4 splitter and the gas plant. For more on FCC operating modes and constraints, see FCC unit equipment layout, operating modes and performance optimization.
Has DLPC® been applied in ethylene plants?
Imubit has worked on optimization strategies for C2 and C3 splitters and for steam cracker furnaces. If you have a unit in mind, contact us to talk it through.
How do you handle multi-objective optimization?
Models can be trained with multiple objectives. Training introduces tunable parameters, so that in real time the user can change the relative importance of each objective.
For an FCC with varied feed composition, does the controller account for it?
Yes, it has to, and it does. A refinery might run three or four base crudes in changing proportions, and the FCC feed changes with them. The key is a training data set with enough operating data in each feed composition regime, so that the controller trained by reinforcement learning knows how to respond across the wide range of scenarios it may meet in real-time operation.
Continuous improvement
How does the application work around instrument limitations, such as a minimally instrumented site or an MV or CV outage?
If an outage is unplanned, the controller switches itself off. For planned outages, contingencies can be built into the model:
- A flexible, or flex, MV covers an instrument on a line that is in service for part of the time.
- A CV can carry a flexible constraint when the unit controls to that constraint some of the time.
- Bad-value protection can be configured, with instructions for what to do when a value falls in a bad range.
How much input does DLPC® need from process and control engineers? Can the models be adjusted?
Process and process control engineers are brought into the modeling to contribute their domain expertise. Deep learning models are not magic, and data alone is not enough. Domain experts define the rules of the problem: which variables are independent and dependent, the economic objective, the time to steady state, the constraints and so on.
The model learns as much as possible from the data. Where important information is not fully contained in the data, the engineers' knowledge is added, for example mass balance or gain tuning. Automated machine learning (AutoML) translates that domain expertise into data science parameters, so the engineers do not need a machine-learning background.
What does it take to retrain the model on a new operating regime, for example after a turnaround changes a unit?
The rate-limiting step is collecting enough data in the new regime to represent a broad spectrum of operation. The retrain itself follows once that data has been collected.
How often are models retrained, and which plant teams are involved?
After the initial training, retrains need little input from the plant. Plant data is collected continuously, so the site does not have to export it. As part of continuous improvement, the Imubit team monitors each DLPC® application and proposes a retrain when it sees a shift in operation. The Imubit team does most of the retrain work and brings in the site team to review the model updates afterward.
Technology comparison
How does DLPC® differ from a traditional APC package?
Of the many technical differences, two matter most: the process model and the solver.
The dynamic process model is a neural network trained on the unit's own operating data, rather than a model built from first principles alone. A data-first model avoids the linearizations, approximations and assumptions that first-principles models depend on, while engineering knowledge such as mass balance is still added where the data falls short.
For the solver, reinforcement learning trains the controller offline in simulation, so the optimization is solved ahead of time across the range of scenarios the unit may meet. There is no online solver calculating each move at run time. A child learning to walk is a useful comparison: the child learns by trial and error and then walks without calculating each step, where an online optimizer measures and calculates every step as it is taken.
DLPC® runs on top of the existing APC or DCS and writes setpoints, inside limits set by plant engineers.
People
How hard is it to bring operations teams on board with AI-based process control?
At one refinery, the operations team was skeptical at first, both about whether the technology would work and about what it would mean for their roles. Starting with a simple initial application helped build operator trust. The site chose a debutanizer tower with two MVs that operators already knew how to move and interpret, so they could see the results and understand them. That early trust matters before moving to more complex control strategies, where explainability gets harder.
