PTQ Q3 2022 Issue

While this method has had success with true linear pro- cesses and managing a high volume of local constraints, maximising the FCC value potential requires discovering all the critical relationships and modelling non-linearity. Leveraging years of the refinery’s FCC historical data, a CLNN process model learns over thousands of historical windows using deep learning AI to model relationships between all key independent and dependent variables through prediction of dependent variables dynamically. Through offline training, the CLNN process model learns that relationships change based on continually changing process factors such as composition shifts and fouling. Instead of a single gain value used to represent the rela- tionship between an independent and a dependent vari- able for all operating conditions, CLNN learn that there are thousands of gains with varying time dynamics per independent-dependent relationship and knows when to apply the right gain to the dynamic conditions of the FCC in a given moment (see Figure 2 ). The Figure 2 gain distribution plot allows plant experts to understand what relationships the CLNN have learned. Plant experts who understand the operating strategy can see the level of non-linearity, both direction and magnitude, for each gain distribution and confirm that what the model has learned can be supported by process reasoning. In addition to learning the relationships between pro- cess signals, the CLNN process model learns inherent relationships that are not directly measured. For the refin - er’s FCC, the catalyst formulation was modified and was not present in the historical dataset used for initial CLNN training. After several months of new catalyst formulation introduction, the CLNN model was retrained to include the new data. After training, the CLNN learned distinct bimodal gain distributions between the reactor overhead temperature and FCC yields (see Figure 3 ). The CLNN process model was able to learn and classify the difference in yield shifts based on catalyst activity without explicit data about the catalyst properties. This is just one example of how the CLNN dynamic process model can infer unmeasured dis- turbances through the use of frequently measured signals. Human-like experience-based controller Once the CLNN process model is developed, the CLNN dynamic controller and optimiser use the process model as a true FCC simulation environment to master the operat- ing strategy while gaining human-like experience through offline reinforcement learning. The CLNN is run through millions of different process situations where the process state, constraint limits, and prices or targets in an economic objective are randomly set. In each unique situation, the CLNN must drive the eco- nomic objective to optimal while respecting operating limits using defined manipulated variables. With changes to prices or targets in the economic objec- tive, CLNN must learn to adapt by changing product distri- bution based on what is optimal while also understanding the cost of key handles to the economic objective. For exam- ple, while the resid feed price is cheap, the corresponding

1000

800

600

400

Time

Dymanic process models

200

0 0.4 0.5 0.6 0.7 0

0.8

∆ dependent

1 unit independent

products produced, typically slurry and light cycle oil, may not be valuable per the given price deck. In addition, resid increases regen dense bed tempera- ture depending on the amount of coke formation, which will reduce cat-to-oil, further impacting yields. By training on millions of unique virtual situations, the CLNN start to develop human-like intuition. This intuition is critical as plant data will never be perfect. Today, domain experts leverage their years of experi- ence running the plant to close gaps in imperfect data; for example, a planning and economics engineer selecting the best LP case or a process control engineer selecting the best process model for APC. Just as domain experts lever- age intuition, the CLNN apply their gained experience in addition to the current plant data as part of their solution, making smooth and continuous decisions that operators can trust. Even when data is not perfect, the CLNN lever- age their mastery of solving the operating strategy to apply a solution driving towards the global optimal. Online performance of CLNN Once domain experts evaluate and confirm that CLNN vir - tual decisions are respecting operating limits while driving to optimum, the CLNN are implemented on-site, where they directly control and optimise the FCC. In the case of the refiner’s FCC, the CLNN manipulate the total feed, ROT, and several feed sources such as resid and hot gasoil from the vacuum tower to manage feed composition while maxi- mising the objective function and respecting key opera- tional constraints continuously minute-by-minute. With CLNN disengaged, the refiner executes the FCC yield improvement operating strategy via the execution cycle following their defined rules, as stated previously. However, with CLNN engaged, the CLNN understand the dynamic relationships of the FCC via continuously updated Figure 2 CLNN gain distribution histogram showing the thousands of dynamic relationships learned during CLNN model training for a single pair of independent-dependent variables

PTQ Q3 2022