AI algorithm helps optimize AGR, and save amine, steam and power

As terrific as AI is expected to be, it’s not a solution in and of itself. Instead, it’s a feature that can assist other solutions, according to Karthik Gopalakrishnan, solutions architect for digital transformation at Yokogawa.

“The power of AI is taking off because it lets users do things they couldn’t do before, including predicting potential failures, and even making control processes more predictable and stable,” says Gopalakrishnan. “This is useful because close-loop, PID controls and APC have limits in response speeds and numbers of variables. However, users can now add generative AI (gen AI) to open systems, and produce results and recommendations in response to operating parameters, though not yet for closed-loop control.”

Gopalakrishnan confirms that gen AI is often accused of being a glorified search engine. However, he adds that Yokogawa’s end users appreciate it because it helps them more thoroughly sort through all their data that they wouldn’t be able to use otherwise, and routinely develop more accurate and actionable results.

Train, evaluate, integrate agents

For instance, Aramco’s Fadhili gas plant in Saudi Arabia recently implemented Yokogawa’s autonomous-control AI agents to increase efficiency. This solution uses coordinated AI agents based on the Factorial Kernel Dynamic Policy Programming (FKDPP) to directly and autonomously control and optimize acid gas removal (AGR) operations at the facility. FKDPP is a reinforcement-learning, AI algorithm jointly developed by Yokogawa and the Nara Institute of Science and Technology (NAIST).

The AI agents were introduced in three phases at Fadhili, progressively optimizing various sections until autonomous control of the core process in AGR unit was achieved. To ensure safety, Yokogawa first simulated the plant to train the agents, and then evaluated their reliability and validity. Subsequently, they were integrated with Yokogawa's Centum VP integrated production control system to use the safety functions of the existing plant (Figure 1).

Yokogawa and Aramco add their project is undergoing a detailed evaluation, but initial results from Fadhili found 10-15% less amine and steam use, around 5% less power use, improved process stability, and decreased manual intervention by operators, despite ambient condition changes.

Summarize like a human

Gopalakrishnan adds, “We worked with one customer on adding gen AI functions, and found the next level is using AI to handle different process hypotheses, such as dealing with shutdown situations. Gen AI summarizes data like a human, and delivers presentable results right away, instead of the two days it usually takes to get one paragraph. Users typically receive research, trends and manual standard operating procedures (SOP), but we say gen AI can provide them, too, and add them to Yokogawa’s Asset Health Insight platform via a gen AI analytics extension.”

Get your subscription to Control's tri-weekly newsletter.

To implement gen AI and achieve similar gains, Gopalakrishnan reports that users must start by checking their data, and not just their process control information, but other structured and unstructured data sources. Structured sources usually include numbers, tables and columns, organized in Excel or other spreadsheets. Instructed sources typically include photographs, videos and other images of gauges and other instruments, which users previously visited in-person to scan or write down values manually.

“Besides accelerating research and tasks, AI can also break down organizational silos by putting data into more suitable contexts, which makes it easier to apply analytics and other software tools,” adds Gopalakrishnan. “What makes gen AI work is good data, access to it, and governance policies and orchestration. Governance enables data to accurately reflect what’s happening in reality. Orchestration is about how users come to perform analytics, and asks how they’re cleansing their data, and are they saving it the right way?”

AI improves brewing-temperature schedule

In addition, Yokogawa and Craft Bank Co. (craftbk.net) in Kyoto Prefecture, Japan, and Yokogawa recently proof tested a manual temperature-setting schedule created by FKDPP algorithm. It was applied to the fermentation process for the brewer’s Bank IPA 1 craft beer.

Previously, fermentation temperature was kept constant, and the brewmaster manually measured sugar content daily, performed sensory evaluations of aroma and taste, and checked for off-flavors. Consequently, Craft and Yokogawa’s proof of concept (PoC) focused on temperature’s effect on fermentation speed. They confirmed that temperature-setting schedule AI algorithm would be valid and effective.

To implement the PoC, a simulation was developed that replicated Craft’s beer production process. Using data from the brewmaster, including stress conditions on the yeast, the AI algorithm developed the temperature-setting schedule for the fermentation tank within the simulation. The brewmaster reviewed the plan’s appropriateness, and followed it to manually set temperatures. Using sensory evaluation during fermentation, it was confirmed that all quality criteria were met.

By manually implementing this temperature-setting schedule developed by FKDPP, Craft reports it shortened its fermentation time by 28% from 336 hours to 240 hours.