Online Exclusive Technical Q&A: AI’s use in chemical plant operations

Hydrocarbon Processing (HP) sat down with Dr. Hiroaki Kanokogi (HK), General Manager, Yokogawa to discuss how artificial intelligence (AI) can and is being used in the hydrocarbon processing industry. Dr. Kanokogi’s organization recently announced that it used AI to autonomously control a chemical plant for 35 consecutive days. The AI used in this control experiment, the Factorial Kernal Dynamic Policy Programming (FKDPP) protocol, was jointly developed by Yokogawa and the Nara Institute of Science and Technology.

 HP: What makes this AI (FKDPP) different from other forms of AI that can be applied in plant operations?

HK: In the industrial AI sector, the vast majority of AI is what we call “problem analysis AI.” This kind of AI analyzes the data that is provided to detect anomalies for predictive maintenance, predict quality or determine the cause of issues. It is generally used to support human decision-making.

In this case with a chemical plant, we are talking about autonomous control AI, which searches for the optimal control model by itself and then implements that. There are several forms of AI for control (TABLE 1); however, based on the analysis of a global survey in February 2022, our organization confirmed that there were no other forms of AI that directly change the manipulative variable in a chemical plant. We are very confident about this. This uniqueness can deliver a great benefit to customers, as this next-generation control technology can control operations that have been beyond the capabilities of existing control methods (PID control/APC) and have up to now necessitated manual operation based on the judgements of plant personnel.

TABLE 1. Primary characteristics of AI used in plant control





Autonomous control

For areas that cannot be automated with existing control methods (PID control/APC), the AI deduces the optimum method for control on its own and has the robustness to autonomously control, to a certain extent, situations that have not yet been encountered.

Based on the control model it learns and deduces, the AI inputs the level of control (manipulative variable) required for each situation.

The benefits of FKDPP are as follows:

(1) Can be applied in situations where control cannot be automated with existing control techniques (PID control and APC), and can handle conflicting targets, such as achieving both high quality and energy savings.
(2) Increases productivity (quality, energy saving, yield, shorter settling time)
(3) Simple (small number of learning trials, no need to import labeled data)
(4) Explainable operation
(5) Same safety as conventional systems (highly robust, can be directly linked to existing integrated production control systems)

Support for areas with automation built-in

AI can take over the task, currently performed by operators, of inputting target values (set value) for areas where automation has been implemented using existing control methods (PID control/APC).

AI uses past control data to perform calculations and enters target values (set value).

・Automation of manual tasks and achievement of stable operations is possible.



Operational support for people

AI proposes target values (set value) that operators will refer to when performing operations.

AI uses past control data to suggest target values (set value) to humans.

・Differences due to operator proficiency level will disappear.

HP: What were the major benefits of incorporating AI within the chemical plant setting? 

HK: It could autonomize an area that could not be automated with existing control methods, while ensuring safety and improving productivity.

Until now, there have been many parts of the plant that have not been fully automated. The next generation control technology using reinforcement learning-based AI (FKDPP) will autonomize areas that could not be automated with existing control methods while ensuring safety and improving productivity. FKDPP is a disruptive innovation that allows for a different dimension of control, particularly in such areas. This AI technology can be applied in the energy, materials, pharmaceuticals, and many other industries where the daily monetary value of operations in large-scale plants is in the range of tens of millions of dollars. Autonomous control AI (FKDPP) can greatly contribute to the autonomization of production around the world, ROI maximization, and environmental sustainability, and will have a major economic impact.

HP. How can FKDPP generate control model in only around 30 learning trials?

HK: Autonomous control is possible with our unique and original algorithm that requires only around 30 learning trials. Yokogawa has been developing the control AI since 2017. Yokogawa’s core competence and strength lies in measurement, control, aggregating information, and producing value. This unique AI algorithm incorporates our operational technology (OT) know-how on the gathering of sensor data from throughout plants to optimize plant operation and control. By implementing the knowledge Yokogawa has for controlling plants, we can eliminate the number of calculations drastically and generate the control model with that number.

There is no AI that is fit for all purposes. Wolpert and other people proved mathematically that "machine learning can produce excellent results when it is domain-specific" in 1995. This is a famous theory to predict the development of AI and machine learning by domain. AI specific to a particular field or domain may exceed human capabilities. So, both deep understanding on AI itself and the domain knowledge are required.

HP.  Can AI do all operations? Or are plant personnel still needed for operations?

HK: In the field trial, the AI directly controlled the operations through the DCS without the need for human intervention. This AI has the potential to be used for controlling a wide range of operations across a variety of industries. However, although the AI can carry out optimal operations of the controlling point, plant personnel are still needed to monitor the status in the control room, just like we do for PID control and APC.

HP. Where can we utilize FKDPP in the energy industry?

HK: There are still many operations in the energy industry that are difficult to control automatically, and so are basically still managed manually by skilled operators. This is because chemical reactions tend to be nonlinear and are affected by disturbances, so that makes it difficult to use a mathematical approach using PID. We think that there is a possibility to enable optimal and autonomous control in such difficult areas.

One example is the control of the large boilers that produce the steam used by rotating turbines for thermal power generation. A related application is gas combustion control in gas turbines. Regarding renewable energy, controlling how geothermal energy is efficiently used for power generation is quite a challenge. FKDPP allows us to leapfrog the automation stage and go directly from manual to autonomous control in these kinds of areas.

HP: What was a major takeaway from this exercise? 

HK: The biggest takeaway was that we can ensure safe autonomous control with AI that improves productivity and reduces cost and time loss.

This test confirmed that reinforcement learning AI can be safely applied in an actual plant and demonstrated that this technology can control operations that have been beyond the capabilities of existing control methods (PID control/APC) and have up to now necessitated the manual operation of control valves based on the judgements of plant personnel. Also, losses in the form of fuel, labor costs, time, etc. that occur due to production of off-spec products were eliminated.

HP: What's next for this form of AI? Do you plan on deploying this on other petrochemical/refining units?

HK: We are certainly looking to work with customers on field trials for other processes and applications to confirm the versatility and robustness of FKDPP, and demonstrate the value in terms of the profitability and sustainability benefits it can deliver. This time, we established and verified the three steps for ensuring safe operations. Next, we need to streamline this process so that customers can test and deploy this technology as quickly as possible.

HP: Going forward, do we foresee AI replacing the traditional method (PID)? Or will it be limited to few niche applications?

HK: FKDPP can be applied to most kinds of control including situations that could not be automated with existing control techniques (PID control, APC). Not only that, we have confirmed in a variety of application experiments that FKDPP can achieve stabilization 1/2 to 1/3 quicker than conventional control (PID control), without overshooting. This characteristic will be beneficial for customers who have furnaces and injection molding machines.



Dr. Hiroaki Kanokogi is the General Manager at Yokogawa. He joined Yokogawa in 2007 and is currently pursuing the development, application, and commercialization of AI designed for production sites. Dr. Kanokogi is one of the inventors of the FKDPP algorithm, and he was previously engaged in machine learning application R&D at Microsoft Japan. He holds a Ph.D. from the University of Tokyo.


Related News

From the Archive



{{ error }}
{{ comment.comment.Name }} • {{ comment.timeAgo }}
{{ comment.comment.Text }}