necessary knowledge to always drive to a global optimum. The simulator has also served as an offline sandbox at the facility, where engineers can run ‘what-if’ scenarios to sim- ulate the outcome of potential operating scenarios before making recommendations to operations. w Reinforcement learning controller The dynamic process simulator is used as the training environment for the application of advanced reinforce- ment learning techniques, which teach the controller how to manipulate key control handles to optimise the process within user-defined constraints. The controller undergoes hundreds of thousands of simulations in this process model, enabling it to learn not only from every operating scenario experienced in the recorded history of the plant, but also from hypothetical scenarios. Once trained, the con- troller takes two forms: the on-premises model connected to site advanced process control (APC) or distributed con- trol systems (DCS), and an offline copy in the cloud. The on-premises model is fully isolated from the cloud training and development environment, ensuring compliance with site safety and cybersecurity standards. The offline model provides another playground for operators and engineers to use the trained AI model to explore different strategies while anticipating the system response to new constraints or conditions. Success Collaboration with the technology partner on application design drew on the expertise of console operators, process engineers, and leadership. This scoping process helped define true operational pain points while fostering own - ership and buy-in. During commissioning, they received hands-on training for operators, focusing on interactions with the application, such as setting constraints and targets effectively and understanding how to interpret controller adjustments. Meanwhile, the engineering team used the cloud-based industrial AI platform for real-time monitoring and ‘what-if’ scenario analysis. Unlike one-time optimisation projects, this AIO solution incorporates ongoing support through routine check-ins with operations, engineering, and leadership teams. This collaborative support process identifies new opportunities, addresses model performance shifts, and maintains align- ment with the refinery’s key performance indicators (KPIs). By embedding governance structures and open feedback channels, Big West has been able to sustain high applica- tion engagement and ensure continuous value generation from the technology. Case studies One of the keys to successful sitewide adoption of AIO was its ability to capture and communicate its positive outcomes. Some of these successes drove curiosity across units at the site through significant operations improve - ments to throughput or quality. Other times, the bene- fits were softer. Operations began to question the status quo and leverage the AI models to answer questions like What if we could push more flow through this valve that’s
always constraining us? What would that do for overall unit throughput and our objective function? With the technology necessary to optimise to the current physical constraints of the unit, operations and engineering were quickly identify- ing the next debottlenecking opportunity. Case study 1: Debutaniser throughput optimisation The FCC debutaniser provides a final separation of the heavier components of the fractionator overhead stream, producing a cat (catalytic cracker) gas stream and an olefin stream. Since cat gas is typically a more valuable product, there is an economic incentive to minimise excess olefin production, beyond what is required to fill the alkylation (alky) unit. Furthermore, cat gas Reid vapour pressure (RVP) minimisation is critical for maximising the volume of butane upgraded into the gasoline pool. Historically, these two incentives drove operators to maximise debutaniser reflux while simultaneously maximising reboiler duty to product specifications or flooding constraints. • Challenge: The undersized debutaniser tower almost constantly ran near or beyond the flood point, leading to difficulty maintaining clean separation and producing on-spec products. The debutaniser bottleneck frequently limited the FCC conversion, and a more conservative oper- ational approach was taken that prioritised keeping the tower out of flood and maintaining stable operation over pushing to constraints. However, this approach lacked both consistency and robustness. Operators frequently adjusted tower operation due to unpredictable flooding, leading to One of the keys to successful sitewide adoption of AIO was its ability to capture and communicate its positive outcomes additional suboptimal time following the event, dialling in the tower to bring products back on spec. There was no consistent framework to address these issues, which led to variability in results depending on the operator running the console at the time. • Solution: Big West had two primary control handles – reboiler duty and the temperature at the reflux return – to address flood control and yield optimisation. Collaboratively with Imubit, the team designed an AIO application for the debutaniser, leveraging the previously mentioned process. It calculated the necessary inferentials, created a process model, and put that model through the rigour of years of simulated experience via reinforcement learning. The result was a closed-loop optimiser that leveraged that experience to control the debutaniser and achieve the desired eco- nomic strategy. Operations engagement was imperative to this model-building process. Senior operators validated model predictions and participated in training sessions, which boosted confidence and buy-in. • Results: As a result, the debutaniser achieved a 2% increase in throughput, mitigating previous FCC conversion
55
PTQ Q3 2025
www.digitalrefining.com
Powered by FlippingBook