FCC unit yield improvement with artificial intelligence
How closed loop neural networks have improved FCC yields via direct control and continuous optimization.
Geraldine Hwang and Abishek Mukund
Viewed : 142
Refineries have 10 or fewer key operating strategies that, when executed efficiently, maximise plant value generation. Yield improvement on a fluidised catalytic cracking (FCC) unit is one of the most profitable operating strategies in a refinery by optimally converting low-value gasoil and sometimes resid into high-value gasoline, olefins, and diesel.
However, this operating strategy is also one of the most complex. The FCC process is highly dynamic, with an intricate energy and mass balance. Dynamics are also impacted by unmeasured disturbances such as catalyst activity and equipment fouling. Key mechanical limits, such as regenerator (‘regen’) temperatures and slide valve differential pressures, and environmental limits, such as regen CO, need to be respected and controlled to ensure unit availability.
In addition, the FCC is heavily integrated into the entire refinery system, facing impacts of feed composition changes from the crude vacuum tower as well as impacting downstream units and integrated utilities. While staying within operating limits and balancing unit dynamics, decisions need to be made to generate the most profitable yields with the most economical handles, such as total feed rate and conversion by manipulating catalyst circulation.
Plant experts understand the complex reality of the bespoke FCC description. However, traditional solutions cannot solve this problem entirely, forcing domain experts to break down the FCC operating strategy into distorted fragments of the true problem.
Common FCC operating strategies are defined based on an economic objective function or pushing to a downstream constraint such as alky feed operating limits. In either form, knowing where to set conversion handles, such as reactor overhead temperature (ROT) and feed rates in the dynamic environment with changing constraints, becomes critical to maximise profit.
However, as domain experts do not have a solution that can take on the complexity of the FCC, experts are forced to break down the problem into components: economics, process, automation, and execution. Economics are managed by planning and economic groups, where they define the optimal targets for key handles such as feed and ROT on a weekly basis after running the linear program (LP) over several cases.
Process engineers focus on unit-specific engineering limits, helping to model unmeasured limits such as catalyst activity and manage local constraints with the use of first- principle simulations. Operators focus on safe and stable operation, using rules of thumb and heuristics to keep units steady while executing operating orders as a secondary objective.
Finally, process control groups focus on automating the execution of operating orders while managing stable control using technologies like advanced process control (APC). Each expert group is solely focused on their fraction of the objective, viewing the FCC through the lens of their siloed traditional approaches. Traditional approaches distort the true problem by imposing theory-based, linear, and stratified time assumptions.
Capturing unrealised value
Experts know that the FCC needs to encompass all these components to drive to optimal, so they attempt to reconnect their sub-problems through communication and manual execution of operating orders in a weekly struggle. However, the cycle always ends with the FCC far from optimal. Plant expert groups waste time and resources, leaving millions of dollars of unrealised value potential.
Proven in industry over the past several years, a specific artificial intelligence (AI) process optimisation solution called closed-loop neural networks (CLNN) has been able to model the entire FCC operating strategy to capture previously unrealised value for refineries. CLNN have three critical differentiators that enable plant experts to optimise operating strategies.
First, CLNN holistically model the true dynamics of the FCC using deep learning trained on years of plant historical data. Second, the CLNN model trains over millions of virtual simulations to master control and optimisation of the FCC, acquiring human-like intuition, which it applies during direct control of the plant once implemented on-site. Finally, the solution allows all plant experts to understand the model, representing a single view of the FCC operating strategy but encompassing each plant group’s perspective.
Applying CLNN to an FCC
A refiner with a 90-105 kbd FCC traditionally optimised their FCC operating strategy based on a weekly execution cycle running the LP solver and process engineering simulations to give operating orders to operations that use APC. It should also be noted that the FCC was fouling at an accelerated rate.
The refiner measures success through an economic objective function, classically calculating the difference between the value of products made minus the value of feed used. Every week, the planning and economics team runs its LP and makes a binary decision: either feed over conversion or conversion over feed.
Based on market economics for the refiner, the LP typically chose feed over conversion. The planning and economics team then defines a total feed rate target for operations that slowly change the upper limit in APC to reach the total feed rate target. The planning and economics team felt it was most profitable to push resid into the FCC.
To the best of their ability, process control engineers define a set of operating rules for the APC to drive towards the operating orders. The rules are as follows:
• Maximise total feed rate to the upper bound
• Maximise resid rate to the regen bed temperature limit
• Use ROT to manage key constraints such as slide valves and wet gas compressor (WGC) limits.
Knowing the APC programmed decisions, the process engineer and operators then change the ROT bounds and feed rate sources setpoints manually to manage constraints, attempting to manage new limits associated with unit fouling and ensuring the APC will maximise feed to the upper bound. While manual execution of simplified rules attempts to execute the operating strategy, it unintentionally destroys value.
Every move made on critical handles on an FCC impacts critical constraints and the economic value by changing yields. This problem gets exponentially complex when multiple handles share common constraints and impact yields. For example, feed rate, ROT, and feed compositions all share relationships with the products made and the WGC suction limit.
In the traditional approach, the LP models relationships between key handles and yields but does not account for constraints at the granularity of the APC. However, the APC does not understand the relationship between key handles and yields, relying on the LP targets and theoretical engineering models to give that information via targets and rules. To capture full value, a dynamic process model must model both yield and operating limit relationships with key handles in a holistic process model (see Figure 1).
Add your rating:
Current Rating: 1