Figure 1. Level controller can’t influence
The initial article in this series, "Neglect Level Control at Your Peril," examined the behavior of level processes (such as in Figure 1). It specifically looked at the behavior of the reset mode when applied to integrating processes as most industrial level loops are. Especially for large vessels, traditional tuning techniques suggest unreasonably large values for the controller gain, Kc. Applying the Ziegler-Nichols equations to the vessel in Figure 1 gives the following results for the gain and reset time, TI:
Kc = 46 %/%
TI = 1.33 min
The usual action is to reduce Kc. However, for an integrating process, the result is a slowly decaying cycle with a long period. The previous article clearly illustrated this for a noise-free environment.
Here, we'll first examine the behavior in a more typical industrial level loop with noise on the level measurement and frequent changes in flows into or out of the vessel. Then, we'll look at surge vessels -- where Kc intentionally is set to as low a value as possible.
You can use large controller gains such as 46 %/% only in loops with virtually no noise, which essentially restricts them to temperature loops. As illustrated in the previous article, using such a Kc in a noisy level process translates the noise in the level measurement to higher amplitude noise in the signal to the final control element.
Because the amplification of the noise primarily stems from the controller gain, the logical action is to reduce Kc. The main consequences are:
• decreased amplification of the noise; and
• sacrifice of control performance, especially in terms of speed of response.
Figure 2 presents the performance for the following tuning: Kc = 2.0 %/% (a modest amplification of the measurement noise); and TI = 1.33 min (the same as suggested by the tuning equations).
The previous article demonstrated this tuning produces a cycle in a noise-free environment. Figure 2 depicts a similar cycle. The variability in feed flow makes the cycle more difficult to quantify. However, especially in the controller output, cycles of a sinusoidal nature are evident in most intervals of the response. The 12-hr period covered in Figure 2 shows at least 26 distinct peaks. This suggests a period (P) of 720/26 = 27.7 min. Arbitrarily selecting two of the peaks gives a period of 18 min. In either case, a reset time of 1.33 min is far too short -- TI should be at least half the period of the cycle.
Visually assessing the nature of cycles in a response is both imprecise and somewhat subjective. Although not routinely applied in the process industries, a more rigorous approach is to apply the Fourier integral to determine the contribution from each frequency to the response. Those loops that exhibit cycles deserve attention. A cycle usually has consequences, so you must understand its origin. Perhaps the cycle stems from the controller or perhaps it originates with a disturbance.