Smart manufacturing for excellence in process plants: Part 2

Greg: Now to get serious. We gain insightful practical guidance from Umesh Mathur, P.E., Houston, Texas, on the use of real-time optimization to achieve excellence in process plant automation.

What is real-time optimization (RTO)?

Umesh: Unfortunately, “optimization” is one of the most misused terms in our industry. Most engineers think improving operations by trial-and-error (i.e., hit-or-miss, heuristic schemes) optimizes their process, but it’s hardly ever true. Each individual attempting process improvement will reach a different result.

However, the term optimization refers to a system in which a first-principles mathematical model is created to accurately describe process behavior. This model should mirror the independent (operator set) and dependent (unit output) variables. For example, reflux flow rate and reboiler heating medium flow in a distillation column are independent variables (IV). Product purities are dependent variables (DV).

For large applications, such models are expressed as an open equation, where all variables are moved to the left side for each model equation, and the right side is identically zero. Large processes can have a million equations.  However, most equations have only a few variables, so the incidence matrix is generally extremely sparse. These modeling systems are also called equation-based models. Open-equation process modeling enables simultaneous solving of all unit operations, including material or energy recycle streams.

Sequential-modular simulators, on the other hand, solve each unit operation using canned convergence methods. Examples include AspenPlusT, HysysT, ProSimT, ChemCadT, ProMaxT, and others. Recycle-stream convergence generally requires sequential runs of the entire model. These simulators are most useful for offline process engineering work. Unfortunately, they’re computationally inefficient and cumbersome in large-scale RTO applications.

Non-linear mathematical, RTO software can be connected to an open-equation model. The software is designed to work with a sparse matrix representation of the process model. The optimizer seeks iteratively to maximize an objective function, such as overall plant profit, by systematically manipulating the IVs. At each iteration, the RTO optimizer checks the model outputs to ensure all DVs (safety, environmental, process, equipment and product-quality constraints) don’t violate their respective upper and lower allowable limits. It adjusts for IVs at each iteration to ensure that DV constraints aren’t violated, while also seeking to maximize profit.

This process stops when all variables are within their limits and no further IV changes will improve profit because the profit has been maximized, and the optimized values of all IVs and DVs are now defined. This is the optimal, feasible solution because all variables are within their respective upper and lower limits, and profit can’t be increased further.

Commercially successful closed-loop, real-time optimization (CLRTO) software examples include AspenTech’s RT-OPT, Aveva’s ROMeo, Honeywell’s Nova and Yokogawa’s Dynamic Real-Time Optimizer (RT-OP). 

CLRTO refers to further refinement where:

• The plant model is first recalibrated against actual plant data whenever steady-state conditions are found to exist;
• The upper and lower limits (set by operations or engineering) for all IVs and DVs are taken from the DCS;
• Economic variables used to define the profit function (unit values of products, feedstocks and utilities) are updated; 
• The optimization is run to maximize profit, while observing all constraint limits for the IVs and DVs; and
• These optimized values for the IVs and DVs are downloaded to the lower-level multivariable controllers, such as AspenTech’s DMC, Honeywell’s RMPCT and Emerson’s DeltaV PredictPro. These controllers drive the process, minute-by-minute, to the optimal condition by manipulating underlying DCS setpoints. All real-time safety, environmental, process, equipment and product quality constraints are observed when computing the future trajectory of the controlled variables.

Accordingly, the term CLRTO is restricted to systems where a rigorous optimization of the first-principles mathematical model is first carried out, and the optimized results are then downloaded automatically to the underlying, multivariable control layer for real-time execution.

Greg: Where have CLRTO applications been deployed successfully?

Umesh: CLRTO applications have been deployed successfully for more than 35 years in petrochemical and refining industries, such as:

• Olefin plants and many downstream petrochemical facilities, and
• Refineries with crude and vacuum units, gas plants, residual fluid catalytic cracking (RFCC) and other conversion processes.

Greg: What are the benefits of CLRTO applications in the process industries?

Umesh: Refinery and petrochemical CLRTO applications have shown documented paybacks from less than a month to two years. It depends how well the underlying layer of multivariable controls was implemented. Long-term success requires a high level of commitment by managers and owners to train and retain engineering staff capable of maintaining or enhancing such systems.

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

Any RTO system that doesn’t connect to an underlying layer of multivariable, model-predictive controllers (MPC) is called an open-loop or advisory system. Open-loop applications hardly ever provide sustained economic benefits. Closing the RTO loop with the MPC layer is, therefore, an essential requirement for CLRTO.

Greg: What are the requirements for ensuring long-term success for CLRTO applications?

Umesh: This is a crucial question, often insufficiently emphasized during initial project approval. Open-equation process modeling for large units requires using proprietary software that lets users:

• Generate model equations automatically;
• Connect models to sparse-matrix CLRTO optimization software: All underlying unit operations must first converge tightly, so the overall, open-equation model (including all recycle streams) can converge successfully;
• Define the economic objective; and
• Define upper and lower limits of all IVs and DVs.

Over time, process parameters, such as heat exchanger fouling coefficients, kinetic terms in reactor models, and compressor and pump efficiencies, can change gradually. Before commencing an optimization run, CLRTO software must detect steady state, and adjust process parameters automatically to match current process conditions (data reconciliation and parameter estimation). These tasks are most conveniently performed using one model that employs an open-equation modeling framework for simulation, data reconciliation/parameter estimation, and economic optimization.

Open-equation modeling and optimization requires a much more user training than sequential-modular simulation software. It’s important for managers and owners to retain staff who have developed systems for any given CLRTO application because modeling changes may be required to reflect plant configurations or equipment changes that may occur over time.  Often, the initial project is implemented by the CLRTO vendor’s staff. Owners are advised to assign in-house process engineering staff to participate in all aspects of project execution, so they can maintain these applications independently.

Greg: How easy is it to maintain CLRTO applications?

Umesh: In general, CLRTO applications are connected to a lower-level MPC that itself is connected to the DCS. Over the years, maintaining CLRTO applications requires close collaboration between the process modeling and optimization team and the MPC team for every RTO application.

It’s important to maintain staff expertise on both teams. The ongoing economic benefits of successful CLRTO/MPC projects are substantial. They justify investments by owners in training and retaining engineering staff to keep applications running reliably in closed-loop mode, even as they’re typically staying online 99% of the time or more.