Executive Viewpoint: Interconnected neural networks drive breakthrough optimization
We have all learned that technology alone will not improve our operations.
We have all learned that technology alone will not improve our operations, it is how we use technology as an organization and align our teams around it. Using artificial intelligence (AI) to improve the continuous optimization of a refinery or chemical plant is no different. Augmenting your current optimization methodologies and procedures with AI will not create a step-change in plant process optimization because the AI is limited by convoluted team structures and siloed modeling tools.
Limitations of the traditional control and optimization hierarchy
The sheer complexity of controlling and optimizing plants gave rise to what we recognize today as the traditional hierarchy of technologies and teams. Under this hierarchy, a narrow top layer—planning and economics—governs the lower layers. Below this planning layer is process engineering, which is responsible for operating areas and units in the plant in accordance with the targets provided by the upper layer. Below the process engineering layer is process control. Oftentimes, this layer is divided into two parts: advanced process control (APC), which includes multivariate model predictive controllers, and a distributed control system (DCS), which regulates valves via proportional integrative derivative (PID) controllers. The bottom layer of the hierarchy is the field instrumentation layer of valves and indicators.
This hierarchy is designed such that the most economically critical decisions are made at the top, typically on a weekly basis. The decisions then cascade down through the layers, eventually controlling on a minute-wise or second-wise basis thousands of process variables in the form of pressure, temperature, flow, level and quality specifications. Process engineering and operations teams are familiar with carefully managing and optimizing these thousands of variables. Keeping each of these variables optimized at all times is collectively important to plant profitability. However, most of them are not individually critical to keep optimized continuously. By contrast, hidden amongst these thousands of variables are 10–15 pivotal process variables that have asymmetric economic importance. If these 10–15 process variables are identified and optimized continuously, refineries could generate $20 MM–$30 MM in additional margin from their most profitable processes.
One example of a pivotal process variable can be found in conversion units. Planning and economics will provide a general target for conversion or reactor temperature, accounting for the feedstock and the product economics. What if the reactor could always run at the optimal temperature, down to 1° of accuracy every hour of every day? What if this temperature continuously adapted to small unmeasured changes in feed composition and unit conditions to avoid under-cracking or over-cracking the molecules? The answer is millions of dollars of incremental annual value in the product pool.
Under the traditional control and optimization hierarchy, the plant has no way of keeping such a conversion variable tightly optimized at all times, because it requires making decisions at a plant-wide scope. Plant-wide decisions are made at the top planning and economic layer on a weekly basis. By the time the economic decision trickles down to the process control layer, the unmeasured feed disturbance will be long gone and the opportunity to convert the feed more accurately is lost.
The conventional hierarchical control and optimization layered structure is designed for “regular” variables, where general economic guidance is satisfactory, and the precise control of the variable can be done at the local level. The conventional hierarchy is not positioned to keep these “special” variables optimized continuously, since this combines the local process scope and the global plant scope. Unfortunately, the current trend of integrating AI into the different layers of the hierarchy can provide some local value, but not fix the fundamental structural problem. This is where the processing industry has a breakthrough opportunity for a process optimization step change.
Thinking beyond the traditional control and optimization hierarchy
Capturing the opportunity behind the asymmetrically important special process variables requires us to change our mental model around the hierarchy of control and optimization. Similarly, operating the thousands of regular variables must be kept intact, along with the layers and teams that control them. However, we must create an interconnection between the special process variables, related process variables in other parts of the plant and higher layers such as process engineering and planning and economics.
This interconnection is not only between models at the different layers of the control and optimization hierarchy, but between people in the different layers. This interconnection should take the form of a single model that aligns the different teams from various disciplines and from different areas of the plant, all centered around the continuous optimization of the special variables. This should all be done side-by-side and without any disruption to the traditional control and optimization hierarchy, keeping the thousands of regular variables intact.
An interconnected neural network to optimize key plant variables
This interconnection is already happening in multiple plants across the industry. It is empowered by neural networks that have revolutionized the way refiners approach and manage optimization for their most complex and profitable processes (FIG. 1). These neural networks can identify the 10–15 special variables that, if optimized continuously, can have the biggest economic impact. These neural networks align people across different teams on the most valuable problems and operate in closed loop to unlock hidden margins across multiple plant process units.
FIG. 1. The interconnected closed-loop neural network integrates operations, processes and variables throughout the traditional hierarchy.
One further substantial advantage of the interconnecting neural network is the ability to compensate for the lack of reliable real-time feedstock compositional data. The neural network can compensate for this typical lack of composition data using pressure, temperature, flow and level indicators that react to feed composition changes through patterns that are discovered by the neural network. Using historical data, the neural network learns subtle nonlinear relationships between the special variables and manipulates them either through APC targets or PID setpoints.
Takeaway
The traditional control and optimization hierarchy, combined with advancements in software technology and tools, has supported increasingly complex operations over the years. However, we are all looking for ways to operate better and rethinking how we are organized to run the plant more optimally. It is imperative to understand the asymmetrical criticality of the continuous optimization of a very small group of special process variables. Through the integration of a closed-loop neural network, refineries and petrochemical plants can interconnect their people and process areas to optimize their most dynamic and profitable processes and capture millions of dollars in annual incremental margin. HP
The Author
Cohen, G. - Imubit, Houston, Texas
Gil Cohen is the CEO and Co-founder of Imubit. He has led Imubit’s product development and market adoption by leading global refining and petrochemical operators. Prior to Imubit, Mr. Cohen founded and led Cigol, a leading provider of data center system-on-chip network security solutions. He has published several academic papers on ML and has more than 20 yr of experience in developing new algorithms and applying them on previously unsolved problems in various industries. Mr. Cohen earned BSc and MSc degrees in electrical engineering, as well as a BSc degree in mathematics, all cum laude, from Ben-Gurion University of the Negev in Israel.
Related Articles
From the Archive