Abstract: Pumps are often operated poorly controlled and manually throttled, which results
in a significantly reduced energy efficiency. We propose an automatic PID controller tuning
approach in combination with a dynamic flow rate estimation algorithm for low speed radial
pumps to replace expensive flow rate measuring equipment. Since our approach is fully
automated, there is minimal additional effort for the customer to implement an automatic
flow rate control.
5230
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
Combining (5), (6) and (7) yields the process model the fastest possible setpoint tracking without overshoot
sketched in Fig. 2. The response of q(t) to a change in (cf. section 4.2). The second objective is the ±45 deg
n(t) can be expressed as first order model according to control (cf. section 4.3), which results in a reasonable
(5) and (6). The response of qest (t) to a variation in q(t) compromise between speed and robustness, since it leads
is expressed by (7) as described above. After linearization to a damping d = 0.707 in closed loop and thus to a low
the relation between n(t) and qest (t) is described by an overshoot and reduced rise time.
aperiodic second order plus deadtime model (SOPDT).
An aperiodic SOPDT model is uniquely determined by 4.1 Characteristic equation
its response to a unit step. The unit step response can
conveniently be stated in terms of the physical parameters Consider the open-loop transfer function of a PID con-
of the pump and process as follows: troller and plant of arbitrary degree and with deadtime
0, t < TD KC KP −TD s
q̂(t) = (8) Go (s) = e , (13)
fq (KP , TD , TL , TQ , t), t ≥ TD D(s)
with where we assume that the slowest plant time constants
1 t−T
− T D
t−T
− T D have been compensated with the controller time con-
fq = KP 1 − TL e L − TQ e Q , (9)
TL − TQ stants. 1 The denominator D(s) is polynomial of arbitrary
where the deadtime TD and the pumping unit time con- degree z ∈ N. It consists of the plant denominator DP (s)
stant TL are known from (1), i.e., the flow rate estimation with coefficients ak , multiplied by the integral part of the
described in section 2.1. Sections 3.3 and 3.4 describe how controller:
to identify the process gain KP and the fluid acceleration z−1
X
time constant TQ from a step response. D(s) = TC s · DP (s) = TC s · ak · sk , a ∈ R+ (14)
k=0
3.3 Identification of KP All poles so ∈ R−
0 of Go (s) are located on the negative real
axis in accordance with the assumptions stated in section
The gain KP in (9) can be calculated from 3.1. The closed-loop characteristic equation reads
qest,end − qest,start C(s) = D(s) + KC · KP · e−TD ·s (15)
KP = (10)
nend − nstart
for qest,start , qest,end, nstart and nend from a step response. with complex s = sR + sI j, KC , KP ∈ R+ and TD ∈ R+
0.
An example for such a step response is shown in Fig. 4. We assume the dominant pair of roots of D(s) remains to
Since the considered process is nonlinear, KP depends be the dominant pair of roots of C(s) for all stabilizing KC .
on the operating point. In order to ensure closed loop Due to the complex exponential function e−TD s in (15),
stability, the calculation of KP has to be performed for C(s) has an infinite number of roots in the complex plane
all typical operating points and the highest KP must be for TD > 0. Arguably, this is the reason why the root locus
selected for controller synthesis. We analyze the behavior method is not popular for deadtime systems. While the
of KP for the particular laboratory test setup at the construction of the whole root locus plot is tedious in this
beginning of section 5. case, we show in sections 4.2 and 4.3 that the calculation
of the aperiodic limit and the ±45 deg criterion is simple.
3.4 Identification of TQ
5231
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
4.3 Control with ±45 deg root location the pump to simulate various pipe friction factors and dis-
turbances. The nominal pump operation with a rotational
The 45 deg case can be described with similar geometric speed nnom = 1500 min−1 leads to a nominal flow rate
considerations as the aperiodic one. A complex conjugate qnom = 4.69 m3 /h (against fully opened control valve).
pair of roots is located on the ±45 deg locus if and only The control algorithms run on a PC with MATLAB that
if −sR = sI for one of the complex conjugate roots. The is connected to the process I/Os via PROFIBUS.
root We claim the gain KP can be assumed to be constant in
s̃ = −σ + σj, σ ∈ R+ 0 (17)
the test setup. This can be seen as follows. The pressures
satisfies this criterion. Substituting s̃ into (15) yields p0 and p2 are both approximately equal to atmospheric
pressure in the laboratory setup and the geodetic height
C(s̃) = ℜ(D(s̃)) + KC · KP · eTD ·σ · cos(TD · σ) (18) difference z2 − z1 = 0.5 m is small, which in particular
+ ℑ(D(s̃)) − KC · KP · e TD ·σ
· sin(TD · σ) j implies p0 ≈ p⋆2 . Since the gain KP describes the static
relationship KP = limt→∞ q(t)/n(t) after a step in n(t),
Solving C(s̃) = 0 is equivalent to solving both ℜ(C(s̃)) = 0 we may set dq/dt = 0 in (3). Using p1 = pD , which holds
and ℑ(C(s̃)) = 0. KP can be removed from these equations according to the assumptions stated in section 3.1, and
by solving both ℜ(C(s̃)) = 0 and ℑ(C(s̃)) = 0 for KP and substituting (6) for p1 into (3) yields
setting the resulting expressions equal. After rearranging s
and dividing by eTD ·σ this yields q(t) 4 · R · A2 · C0
KP = lim = .
t→∞ n(t) ρ · L · λf − 4 · R · A2 · C2 · n2nom
ℜ(D(s̃)) · sin(TD · σ) + ℑ(D(s̃)) · cos(TD · σ) = 0 (19) The r.h.s. is independent of the point of operation. Figure
The solutions of (19) represent all σ such that s̃ defined in 3 corroborates that a constant gain KP does indeed hold
(17) satisfies the 45 deg criterion. Since σ is equal to the for a large range of operating points.
distance from s̃ to the imaginary axis, we are looking for
the smallest σ ∗ ∈ R+ 0 that solves (19). The corresponding
est
5
controller gain KC (σ ∗ ) can be calculated by substituting
σ ∗ into either the real or the imaginary part of (18). qest(n)
4
[m3/h]
Decomposing D(s̃) from (19) into its real and imaginary tangent
part yields 3
est
flow rate q
⌊z/4⌋
X 2
ℜ(D(s̃)) = a(4k) (−4)k · σ (4k) (20)
k=1 1
⌈z/2⌉
X
− a(2k−1) (−1)⌊k/2⌋ · 2(k−1) · σ (2k−1) 0
k=1
600 700 800 900 1000 1100 1200 1300 1400 1500
rot. speed n [min-1 ]
and
⌊(z+2)/4⌋ Fig. 3. Static relation between qest and n for a geodetic
X (−4)k height difference z2 − z1 = 0.5 m.
ℑ(D(s̃)) = a(4k−2) · σ (4k−2) (21)
2
k=1
⌈z/2⌉ 5.1 System identification
X
⌊(k−1)/2⌋ (k−1) (2k−1)
+ a(2k−1) (−1) ·2 ·σ , Applying the flow rate estimation algorithm results in
k=1
TD = 0.75 and TL = 1.8 for the process model (12).
where ⌈·⌉ and ⌊·⌋ denote the ceiling and floor functions that A rotational speed step from nstart = 1000min−1 to
round their argument to the next higher or lower integer, nend = 1100min−1 is performed to determine KP and TQ .
respectively. Both (20) and (21) can easily be generated The resulting time series of qest (t) is plotted in the left
automatically. Figure 5 shows an example root locus plot diagram in Fig. 4, where the speed step is performed at
along with a plot of C ∗ (s̃) and KC (σ). t = 200 s. KP can be calculated from the time series of
qest (t) using (10). TQ results from minimizing (11) with
5. PERFORMANCE EVALUATION ON A known parameters KP , TL and TD . The right diagram in
HYDRAULIC TEST SETUP Fig. 4 shows the mean squared error, i.e., the cost function
of (11), with its minimum at TQ = 0.55. In summary, the
The proposed controller synthesis approach is applied to process model (12) becomes
a hydraulic test setup that features a KSB Etanorm cen- 0.00449 · e−0.75s
trifugal pump (G32-125.1) driven by a 0.55 kW induction GP (s) = .
(1.8s + 1)(0.55s + 1)
motor with frequency converter, which measures the ef-
fective motor current imeas (t). The rotational speed n(t) 5.2 Controller synthesis
can be varied between nmin = 600 min−1 and nmax =
1500 min−1 . The pump is connected to a piping system We choose a standard PI-controller
with a total length of L = 11 m and an average radius of KC · (TC · s + 1)
R = 0.0125 m. A control valve is placed downstream of GC (s) =
TC · s
5232
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
Step response of qest(t) and q^(t) Mean squared error mse(TQ) over TQ Root Locus KC
3.1 1 1 220
mse(TQ)
200
min(mse(TQ))
3 0.8
qest(t) fq discontinuity
qest(t) / q^(t) [m3/h]
0.5
2.9 qest,start 0.6
mse(TQ)
154.1
qest,end
Imaginary axis
2.8 0.4
q^(t) 120
0
102.9
2.7 0.2
80
2.6 0
180 200 220 240 0 0.55 1 1.8 2 -0.5
time t [s] TQ 40
KC (s)KP
2
80
then becomes: KC (sR) 60
1.5
KC · 0.00449 · e−0.75s 1
Go (s) = (22) 40 0.692
0.99s2 + 1.8s
20
For the aperiodic limit control the maximum of 0
0
0.99 · s2r + 1.8sr
KC (sR ) = − (23) -20 -1
0.00449e−0.75·sR -1.8 -0.629 0 0 0.529 1 1.5
has to be determined. This maximum is located between real axis sR σ
the poles s1 = 0 and s2 = −1.81, where KC (sR ) crosses
zero. The lower left diagram in Fig. 5 shows the plot Fig. 5. Root locus plot, KC (sR ) and KC (σ) · KP function
of (23), with a unique maximum at s∗R = −0.629 and of the sample process (22)
controller gain KC (s∗R ) = 102.9. The root locus plot in
the upper diagram in Fig. 5 shows the branch point of the Estimated flow rates
two dominant roots located at s∗R for gain KC = 102.9. 4
to the deadtime. 3
3
For the ±45 deg control the first zero crossing of C ∗ (s̃) (19) 2 setpoint
has to be detected. Using (20) and (21), C ∗ (s̃) becomes: aperiodic
C ∗ (σ) = (1.98σ 2 − 1.8σ) · cos(TD · σ) + 1.8σ · sin(TD · σ) 1 45 deg
20 30 40 50 60 70 80 90 100 110 120
CHR
The lower right diagram in Fig. 5 shows a plot of C ∗ (σ) Time t [s]
ZN
with a zero crossing at σ ∗ = 0.529. The corresponding 1500
Rotational speeds
KC (σ ∗ ) = = 154.1, (24)
KP · eTD ·σ∗ · sin(TD · σ ∗ ) 1000
5233
19th IFAC World Congress
Cape Town, South Africa. August 24-29, 2014
small overshoot visible on the rotational speed. The ZN crucial if the acceptance level of automatic control is to be
tuning leads to aggressive control action, which amplifies raised. Although the proposed tuning method requires a
the measurement noise, particularly after the disturbance process model, the implementation effort is as small as for
at t = 80s. Obviously, the ZN tuning results in the lowest empirical tuning rules, since the process identification and
robustness and a minor increase in the process gain will the controller tuning are fully automated. Unlike empirical
lead to an unstable closed loop control. We define two error tuning rules, our tuning method guarantees a predefined
measures to assess the control quality quantitatively: closed loop behavior for delayed systems of arbitrary
τ
1 X |qest (k) − qmeas (k)| 1X
τ degree, even for systems with dominant deadtime. The
MREE = ; MSCE = ε(k)2 , proposed approach proved very useful on a real pumping
τ qnom τ system.
k=1 k=1
where τ is the number of measured samples and ε is
the control error. The mean squared control error MSCE REFERENCES
and the measured energy consumption of the pump are
T. Ahonen, J. Tamminen, J. Ahola, and J. Kestila.
listed in the following table, where we normalize w.r.t.
Frequency-converter-based hybrid estimation method
the respective maximum. As expected from Fig. 6 the ZN
for the centrifugal pump operational state. IEEE Trans-
Controller setting MSCE Energy consumption actions on Industrial Electronics 59, no. 12 (2012): 4803-
Aperiodic root locus 79.84% 94.63% 4809.
±45 deg root locus 67.74% 94.78% A.T. de Almeida, P. Fonseca, H. Falkner, and P. Bertoldi.
CHR 100% 95.95% Market transformation of energy-efficient motor tech-
ZN 59.9% 100% nologies in the EU. Energy Policy 31, no. 6 (2003):
563-575.
tuning shows the lowest MSCE together with the highest C. F. Colebrook. Turbulent flow in pipes, with particular
energy consumption due to the continued acceleration reference to the transition region between the smooth
and deceleration of the pump. Out of settings 1 to 3 and the rough pipes. J. Inst. Civ. Eng. London, volume
the ±45 deg setting shows the lowest MSCE while all 11: 133-156.
three settings have similar energy consumptions that are J. Cvejn. PI/PID Controller Design for FOPDT Plants
reduced by approx. 5% w.r.t. the ZN tuning. In order to Based on the Modulus Optimum Criterion. 12th Inter-
evaluate the control quality with respect to the estimation national Carpathian Control Conference (ICCC) (2011):
quality we compare qmeas (t) and qest (t) by using the mean, 60-65.
relative estimation error MREE. The results are shown in H. Darcy. Experimental research on the flow of water
Fig. 7 for all four controller tunings. in pipes (in French). Comptes rendus des seances de
l’Academie des Sciences, volume 38: 1109-1121.
Aperiodic: MREE= 0.019 45deg: MREE= 0.019
4 4 F. J. T. E. Ferreira, J. A. C. Fong, and A. T. de Almeida.
qest(t) / qmeas(t) [m /h]
6. CONCLUSION
5234