Computer-based process simulations have been in routine use for more than 30 years for the design of new plant equipment and also for training. But they are rarely used for the design of control schemes or for the tuning and testing of controllers.
It is not always possible or useful to perform all the necessary studies and tests for tuning and testing on a live process. Therefore, simulations could be particularly handy in these cases.
Process designers were among the first intensive users of these tools. With them, process designers can carry out their work much faster and more easily, as well as study many more different situations and cases. However, these simulations in most cases still are static and not dynamic.
More recently, increasingly dynamic simulators have been developed, mainly for pre-startup training of operators when new equipment or even an entire new plant is installed. Some of them are quite complex and expensive, up to several million dollars. But they easily pay back the investment by enabling the smoother start up of new equipment or a plant, or to bring them to full capacity more quickly.
Here and there, engineers also tried to use such simulators for the development of new controls. In principle, they should offer the same benefits they do in design and training, namely the ability to look at more different scenarios and solutions or simply to save time. But so far they have not become a standard tool for control engineers.
One reason is that until a few years back there were basically only two types of dynamic simulation tools around. Some were large packages that allowed very realistic models but were much too complex, expensive, and difficult to be used in daily life. At the other end of the spectrum were numerous tools that were nearly academic in their simplicity. They were easy to use but they lacked the needed coverage of real-life situations, such as sticking valves, etc.
Recently, however, several more compact and relatively inexpensive packages have arrived on the market that show great promise. Let us take a look at three quite different but reasonably typical cases to see if we could use such tools today in the context of process control.
Distilling Trainer Knowledge
Some time back, I was called by an oil refinery to help develop several product quality controls for the distillation towers of its delayed coker unit. It quickly became apparent that because of the process dynamics the standard PID controller was not usable, and that model-based predictive control had to be applied. As the name implies, this technology is based on a mathematical description of the process behavior. It requires therefore quantitative knowledge of the static and dynamic effects as represented, for example, by the process parameters.
Normally, it is not too difficult to carry out the necessary tests on a distillation tower. But here the situation was different. The unit used several drums to store the coke. Whenever one particular drum was full, the coke was routed into the next one, and so on. These switches happened every few hours and caused a tremendous upset for the plant because the new drum had to be made oxygen-free with steam, heated up, etc.
The consequence of this situation was that the time span during which the plant was running smooth and steady was never long enough to conduct a meaningful test. This of course created a dilemma, because we definitely needed the process parameters.
In checking out the alternatives, we found the refinery still had in the training center a well-kept dynamic simulator, which had been used for pre-startup training of the operators. We hoped we could use it for our test work. Some checks proved that its behavior was realistic enough. Luckily, this simulation even gave us the possibility to disable the coke switching procedure so we could study the behavior of the distillation towers. In the end, all the necessary test were done on the simulator, all the needed information was obtained in a short time, and the controls could be developed and tuned. The simulator helped to overcome an unusual problem and allowed us to finally make a significant improvement in the operation of the unit.
Unfortunately, in other cases training simulators turned out to be not usable. The main reason was, after the end of the pre-startup operator training, there was no person assigned to keep them up to date. They were not a credible representation of the process any more. A real pity, because they were not only useless for control purposes but also for training of new operators. The lesson is: If such a simulator is built, then it should be properly maintained and also used actively for control development.
'Simple' Level Control
The second example has to do with a much less complicated situation. Level control is in principle quite easy and simple, but level problems can have significant negative effects on other, downstream parts of the plant.
In almost all cases, the standard PI controller is used for level control. Normally, the tuning is not calculated based on the process parameters with some suitable methods but simply found by trial and error.
This implies there are several tests with the controller necessary to find the proper settings. These tests are almost exclusively setpoint step-tests, because this is the only test that can be applied under normal conditions. It is certainly not acceptable to create a disturbance in the plant just to tune a single, simple PI controller.
But these setpoint tests do not give us the correct picture, because the setpoint of a level controller is practically never changed. The task of the controller is to deal with disturbances, a situation that typically cannot be tested during the tuning effort. We are thus tuning the controller for the wrong task.
On the other hand, it is simple to simulate the dynamic effects of level in a drum or basin. In the case of a vertical drum, only the diameter and the distance between the measurement taps are needed in order to describe the situation sufficiently. If there is a suitable simulation tool, such as Matlab, MatrixX, etc., available, then the controller can be very quickly tested under the correct situation. If tools such as TOPAS are available, different disturbances can be simulated.
Figure 1: Find Level Best
With such aids, it is also possible to calculate the tuning swiftly and specifically for the prevailing situation. For example, in a case where smooth action on the manipulated flow is needed, we can quickly test if a so-called error-squared PID controller could deliver better results (Figure 1). For new designs, within a few minutes it could be checked to see if the dimensions of the vessel are sufficient for proper control without negative influences elsewhere.
Decision Feedforward
The third case involves a commonly found situation in which we have to make a decision regarding the control scheme structure and are looking for extra information to support that decision. Let us take a simple example: a furnace.
The key control objective is to keep the product temperature as close as possible to the setpoint. If this furnace is subject to frequent changes in the product flow rate, then it is certainly difficult if not impossible to avoid large fluctuations in the temperature just by use of feedback control. We can use a disturbance compensator, a feedforward, to reduce the effect of disturbance and consequently these variations. To do so, we need to be able to recognize the change in the disturbance variable and to react in time.
In the perfect case, all process parameters used in the feedforward are 100% correct and there is sufficient time to react--i.e., the deadtime of the disturbance is longer than the deadtime of the manipulated variable. In this perfect scenario, we could even achieve total compensation of the effect of the disturbance.
But what usually happens is that the deadtime of the disturbance, the product flow in our example, is shorter than the deadtime of the manipulated variable"for example, the fuel gas flow. The only thing we know is that even in the academic case, where all other parameters are 100% exact, we could never compensate the disturbance in full. The feedforward will always act too late.
Figure 2: Feedforward or Not?
This leads to a simple but crucial question: In such a situation, does a feedforward make sense at all? Can it still improve the performace of the temperature or not? After all, it takes extra effort to develop and to maintain it. The answer can be found by calculations, but this is quite a tedious task. Much faster and more convincing is to simulate the situation–once with feedback control alone and once with the imperfect feedforward (Figure 2).
Such a comparison delivers the answer with ease and in very short time. We have to invest some effort though. We have to get at least a reasonable estimate of the process parameters in order to conduct a meaningful study. But once we have them, the simulation is done quite easily and quickly.
In addition, the effort to get the process parameters–at least those for the manipulated variable-- is not wasted, even if we decide against the feedforward. They can be used to calculate the best suited tuning of the feedback controller.
So computer simulations are indeed valuable tools. They're not just for the design of new equipment and for training and practicing, but they can give very valuable support in dealing with common control problems as well as special situations.
However, they are not used as widely as they could and should be. One reason is the persistent misconception that such simulations must be of extremely high fidelity in order to deliver meaningful results. That's usually taken to mean complexity, which in turn often means unacceptably high cost and difficulty in use.
At least the last two examples show this is not necessarily the case; that even relatively simple simulation tools can make valuable contributions toward operational improvement. They can help produce better control solutions, often with less disturbance of the process, and especially can help save time.
Simulation tools are sufficiently realistic. They allow us to describe and study many real-life situations, especially disturbances and problems, in an easy-to-handle way. They can provide all the needed control functionality, and allow also the easy transfer of the results into the DCS.
Hans H. Eder is president of ACT. E-mail him at [email protected].
Simulation of a furnace shows that when a disturbance such as an increased product inlet temperature is introduced (yellow line), adding feedforward (right) can reduce product temperature excursions (blue line) significantly compared to using feedback alone (left).
In a case where smooth action on the manipulated flow is needed, simulation shows the output of an error-squared PID controller (red line) is much smoother than a typical standard level controller.
