During my career, modeling has been an essential part of developing a better understanding of opportunities and a better control system. I have built simulations in IBM’s Continuous Simulation Modeling Program (CSMP), Raytheon’s Advanced Continuous Simulation Language (ACSL), Microsoft’s Visual Basic, Hyprotech’s HYSYS, Emerson’s DeltaV, and MYNAH Technologies’ MiMiC. most cases I focused on the unit operations where control system innovation was possible and beneficial. I tackled all types of unit operations except for distillation. I avoided this unit operation because my associate Terry Tolliver was the world’s best in simulating and controlling all types of columns. My initial applications were in the modeling and control of difficult compressor surge control, furnace pH control, and strong acid-base pH control. In these applications the process response was so fast and/or extreme that simulations were the only way to safely analyze the response and test a solution.
One compressor I modeled could accelerate to the point of rotor damage within a few seconds once in surge. Because the compressor was responsible for all of an intermediate’s production and rotors were one of a kind, the compressor could not be allowed to go into surge. A derivative module was installed that detected a high rate of change of speed and shut down the compressor before axial thrust was even measured. Simulation was the key to understanding how compressor surge was an incredibly fast phenomenon, like falling off a cliff. Compressors in two other applications were not as difficult to model but operations would often trigger a surge and have to shut down the compressors that were being manually brought online. In both cases, Operations insisted that automatic start-up was not possible because manual start-up was so difficult and unpredictable. I prototyped and demonstrated by dynamic simulation the automated start-up of the compressor flow, pressure, speed, and surge controllers. The go-ahead was given and the control systems thereafter provided faster and much more consistent start-ups with no surge. For difficult control problems, I built dynamic first principle models that integrated ordinary differential equations (ODE) for charge, component, energy, material, and momentum balances as documented in Appendix F. Today, for common unit operations you can graphically drag and drop in modeling objects instead of writing and programming the ODE.
Via a test, an auto or adaptive tuner, model predictive controller, or rapid modeler module can identify the dynamics for a given combination of a process input and output. The test typically consists of a momentary step change in a flow, often by a step change in a PID setpoint or output. The identified deadtime, process gain, and process time constant can be used to provide an experimental model. This model is linear and may not reflect all the interactions and effects of operating conditions, but it may be more accurate than the first principle model for operation at the test conditions.
Concept: Experimentation on actual processes is limited at best. Changes in operating points often require documentation and approval. New control systems must improve and not hinder plant performance from the “get-go.” First principle and experimental models can provide a virtual plant to develop and demonstrate new strategies and provide implementation details and expected benefits.
Details: Building and using first principle models based on the conservation of energy, mass, and momentum will help you to develop a deeper understanding of process interactions and dynamics. Process Flow Diagram (PFD) types of dynamic simulators excel at capturing the complexities of process relationships and are useful for detailing the effect of process conditions on process gains. Advanced Modeling Objects (AMO) are more easily integrated than PFD simulations into a virtual plant and offer more realistic dynamics by the inclusion process deadtime, batch and startup responses, and automation system responses. Experimental models obtained from auto and adaptive tuners, model predictive controllers, and rapid modeler modules offer accurate process dynamics for the test conditions. See the ISA Interchange posts What are the types of Dynamic Simulators and their pros and cons? and What is the best approach in developing a dynamic process model? for more details on the relative merits of simulators, fidelity levels, and the steps for constructing first principle models for the Virtual Plant described in Tip #99 and the “Checklist for a Virtual Plant” in Appendix C.
Watch-outs: Steady-state models are only useful for finding gains of self-regulating processes. Most PFD simulators do not extend equilibrium relationships to the driving force equations needed for dynamics. PFD dynamic simulators are notorious for having insufficient deadtime in the model due to the lack of transportation delays, mixing delays, valve backlash and stiction, sensor lags, measurement update times, and signal filters. Nearly all first principle simulators treat volumes as completely mixed and ignore injection delays from dip tubes. For volumes with less than perfect mixing and small reagent or reactant flows, the process deadtime is an order of magnitude or more too low. For fast processes such as liquid pressure and flow, modeling transmitter damping, scan rates, and PID execution times is crucial for getting the dynamics right.
Experimental models can break down due to process interactions and nonlinearities. Experimental models are generally not valid for start-up, shutdown, and abnormal conditions. Backlash and stiction must be modeled to show limit cycles and loop deadtime as the PID output signal works through the deadband, resolution, and threshold sensitivity of the control valve. The PID algorithm in a distributed control system (DCS) is sophisticated and proprietary, representing many engineering years of effort. The PID emulated in dynamic simulators generally does not capture all of the features of the DCS PID. In particular, external-reset feedback, a powerful tool, must be included. For these and many other reasons, the use of the actual DCS configuration in a virtual plant environment (Tip #99) is essential.
Exceptions: Simulations involving computational fluid dynamics and partial differential equations are presently beyond the capability of the automation engineer.
Insight: First principle models and experimental models used in a virtual plant offer a powerful tool for process control improvement that can be extended into an operator training system.
Rule of Thumb: Develop a dynamic model for unit operations where process control improvements can translate into significant plant performance benefits.