A L U M I N I U M S M E LT I N G I N D U S T R Y

Testing cell controller algorithms using a dynamic cell simulator

M. Dupuis, GeniSim Inc.

The two main tasks of an aluminium reduction release. The level of the added noise is func- Mean Square (RMS) noise filtration algorithms
cell controller are firstly to collect and process tion of the ACD, the thickness of the metal [3]. Fig. 1 shows the comparison between the
the raw cell amperage and voltage, and then pad, the amount of sludge and the fraction of noise-free and the noisy evolution of the cell
secondly to use that information to manage the anode surface covered by frozen bath. This pseudo-resistance, which is the slope of the
the cell efficiency. For this it must send instruc- noise level, which can be made to affect cur- voltage/current curve. The purpose of the
tions both to the point breaker feeder in order rent efficiency, can be reduced by automated noise filtration algorithm is to allow the cell
to control the dissolved alumina concentration voltage treatment since version 1.4 issued in controller to use the noisy data so as to esti-
in the bath, and to the anode beam to control 1999. mate the evolution of the slope of the noise-
the Anode Cathode Distance (ACD). The cell controller cannot directly use the free curve. This slope can serve to estimate
There is obviously a major advantage to be noisy cell voltage to calculate the slope of the percent dissolved alumina and so to control
able to test a modification to the cell control- cell resistance, since this would lead to use- the feeding rate. Fig. 2 compares the target
ler algorithms using a simulated cell instead of less results. Since version 13.0 issued in 2011, noise-free slope evolution with three estimates
putting real cells at risk. This is true as long as Dyna/Marc offers linear and quadratic Root of the slope evolution estimated using three
the behaviour of the simulated cell is realistic
enough to provide reliable feedback. In order
to achieve that goal, the Dyna/Marc cell sim-
ulator has been continuously improved since Fig. 1
1994. It has already demonstrated its ability to
reproduce measured cell dynamic evolution,
as shown in previous publications [1, 2].

Testing cell voltage noise

filtration algorithms

Since version 1.0 issued in 1998, the Dyna/

Marc has offered the option of adding noise to
the amperage and voltage tracks in the simula-
tion. For the cell voltage that noise is an out-
put from the simulation. Thus at the end of
Images: GeniSim

each time step noise is added to the calculated

noise free-voltage to present disturbances by
the bath-metal interface motion and by bubble

50 ALUMINIUM · 1-2/2012
SPECIAL A L U M I N I U M S M E LT I N G I N D U S T R Y

different mathematical modes of noise filtra- of the slope of the cell pseudo-resistance in ting with 60 datapoints, each datapoint being
tion applied to Fig. 1. blue. It also presents the estimated slope evo- the results of 5 seconds cell pseudo-resistance
In Fig. 2 the first curve on the left results lution that results from using linear RMS fit- evolution averages. At that time scale, the 2.5
from linear RMS fitting using 60 datapoints
that are themselves each an averaged value
of the raw cell voltage measured over 5 sec-
onds at a 10 Hz frequency. As can be seen,
the resulting estimation is still a bit noisy. The Fig. 2
second curve in the middle results from us-
ing 120 datapoints instead of 60 datapoints.
The result is almost noise-free, but now the
estimation is lagging 5 minutes behind the
noise-free target slope that is being estimated.
This is to be expected, as it is the best linear
fit of cell voltage evolution using the last 10
minutes of datapoints collected, so it best rep-
resents the state of the slope 5 minutes ago.
In the example presented in Fig. 2, the slope
doubles in 5 minutes during a no-feed obser-
vation, so the estimated value is noise-free,
but it is about half of the real value. The third
mode of filtration on the right of Fig. 2 shows
the result of quadratic RMS fitting, also using
120 datapoints. Quadratic RMS fitting of the
cell voltage evolution eliminates the drag in
the slope estimation, which is important, but Fig. 3

for the same number of datapoints used, RMS

generates a more noisy estimation.

Testing feed control algorithms

These days, the majority of algorithms used

to control alumina feed in aluminium reduc-
tion cell are based on continuous tracking or
else on underfeeding and overfeeding cycles,
where the shift from underfeeding to over-
feeding is dictated by a trigger value, which is
based on either the slope of the cell pseudo-
resistance or the slope of the cell normalised
voltage. One of the earliest versions of that
algorithm can be found in Fig. 3 [4]. That
algorithm is available in Dyna/Marc simula-
tor under the name Pechiney Tracking Feed
Control [5].
The basic concept that lead to the devel- Fig. 4
opment of that algorithm was the observation
that the cell current efficiency is maximised
by operating very lean in alumina, and so
very close to the anode effect conditions. The
algorithm then takes advantage of the fact
that during underfeeding, the slope of the cell
pseudo-resistance starts to rise significantly
before the anode affect. Fig. 3 shows the re-
sults by running that feed control algorithm
in Dyna/Marc. The top graphic shows the 24
hours evolution of the cell pseudo-resistance.
Metal is tapped out at noon and anodes are
changed at 18 hours. It can be seen that the
cell is more noisy after the anode change. The
middle graph shows the noise-free evolution

ALUMINIUM · 1-2/2012 51
A L U M I N I U M S M E LT I N G I N D U S T R Y

minutes delay of the noise-free pseudo-resist- but it does affect the timing of the feeding re- decision. The underfeeding rate is 70% of the
ance evolution compared to the estimated gime shift. The third, lowest graphic shows the nominal feeding rate while the overfeeding
pseudo-resistance evolution is not noticeable, feeding periods resulting from the algorithm rate is 140% of the nominal feeding rate. The
overfeeding rate duration was set to 1 hour.
As a result, the resulting evolution of the dis-
solved alumina concentration in the bath in the
Fig. 5 same graph varies from around 2% to around
2.5%, 2% being the alumina concentration
that would trigger an anode effect.
It is important to notice when the feeding
rate is increased that the alumina concentra-
tion continues to decrease by about 0.1% be-
fore starting to increase because the alumina
takes time to dissolve. That delayed response
will trigger an anode effect if the increase in
feeding regime is done too late; hence the im-
portance of eliminating as far as possible the
delay in the pseudo-resistance slope estima-
tion. Fig. 4 shows the resulting 24-hour aver-
aged specific power consumption and current
efficiency: 12.96 kWh/kg and 94.71% respec-
tively.
It is now well recognised that this type of
continuous tracking feed control algorithm
achieves significantly increased current ef-
Fig. 6 ficiency over compared with feed control
algorithms which used a steady feeding rate
most of the time. It is also well known that
the shorter feeding cycle also increases cur-
rent efficiency; this can be tested using the cell
simulator. Fig. 5 shows results obtained using a
shorter, 40 minutes overfeeding rate duration.
As a result, the dissolved alumina concentra-
tion only varies from around 2% to around
2.3%. This leads to a predicted improvement
of the current efficiency from 94.71% to
94.78%, and a slight increase of the specific
power consumption to 13.01 kWh/kg if the
ACD is kept constant.
The demand feed control algorithm devel-
oped by Kaiser and implemented in Celtrol cell
controller [6] is also available in Dyna/Marc.
The same reduction of the feeding cycle study
presented above can be repeated, this time us-
Fig. 7 ing the demand feed control algorithm. Fig. 6
presents the base case results: 12.91 kWh/kg
and 94.67% current efficiency, while Fig. 7
presents results for the case with shorter feed
cycles: 13.09 kWh/kg and 94.65% current ef-
ficiency. Despite a very similar increase of the
feed cycles and reduction of the range varia-
tion of the dissolved alumina concentration,
the two algorithms predict different results on
the global process efficiency: the current effi-
ciency is hardly affected and the specific pow-
er consumption increases more. The reason for
these differences is that with Kaiser algorithm,
it was not possible to maintain the same aver-
age ACD and operating temperature, which
both increased for the shorter cycles case.

52 ALUMINIUM · 1-2/2012
A L U M I N I U M S M E LT I N G I N D U S T R Y

Developing and testing Fig. 10 presents the results of a second run, was removed in the previous run in order to
new feed control algorithms calling for a track every 12 hours only, this keep things more simple. The corresponding
time with the normal anode change event that 24-hour averaged specific power consumption
A dynamic cell simulator can be even more
useful for developing a completely new feed
control algorithm without putting real cells at
risk. One such innovative new feed control al- Fig. 8
gorithm was recently tested using Dyna/Marc
cell simulator: it is the in situ feed control al-
gorithm [3, 7-9].
The main innovation of the new in situ feed
control algorithm is that it can indirectly meas-
ure the concentration of dissolved alumina in
the bath during a no feed track. It does this
by numerically establishing the relationship
between the alumina concentration and the
slope of the normalised cell voltage. In fact,
that correlation is implicitly used in all con-
tinuous tracking algorithms that monitor the
slope of the pseudo-resistance (or the slope of
the normalised cell voltage) to decide when
it is time to shift from underfeeding to over-
feeding.
The cell simulator can quite easily verify
that there is a unique correlation between the
concentration of dissolved alumina in the bath Fig. 9
and the slope of the normalised cell voltage,
and can numerically establish that unique cor-
relation if it exits. Fig. 8 presents the results
by running the in situ feed control algorithm
in Dyna/Marc for 24 hours. A no feed-track is
called every 3 hours in order to evaluate the
dissolved alumina concentration. Fig. 9 shows
the correlation between the slope of the nor-
malised cell voltage and the falling dissolved
alumina concentration. At the top of the hys-
teresis loop, the black line is the fit of the aver-
age path during the tracking, and all 8 tracks
follow the same trajectory. This is why the in
situ feed algorithm can use the equation shown
to establish the alumina concentration at the
end of each track. So there is a unique corre-
lation, because each track starts from identi-
cal conditions, the conditions the in situ feed
algorithm is trying to maintain. Fig. 10
The second innovation at the core of the in
situ feed control algorithm is its use of the pri-
mary calibration surface [3], at the end of each
track, to estimate the ACD after it has estimat-
ed the dissolved alumina concentration. Then,
based on an estimated evolution rate of the
ACD, that same primary calibration surface
is used, together with an assumed ACD value,
to estimate every 5 minutes the dissolved alu-
mina concentration from the cell normalised
voltage. Finally, a simple PID controller serves
to maintain the estimated dissolved alumina
concentration on its target value. In the exam-
ple shown in Fig. 8, that target concentration
was set to 2.25%.

54 ALUMINIUM · 1-2/2012
SPECIAL A L U M I N I U M S M E LT I N G I N D U S T R Y

and current efficiency are: 13.02 kWh/kg and References [9] M. C. Schneller, ‘A new approach to alumina
94.77% respectively. Those results are quite feed control’, ALUMINIUM, 86 (2010) 9, 98-102.
[1] I. Tabsh and M. Dupuis, ‘Process Simulation of [10] M. Dupuis and V. Bojarevis, ‘Retrofit of a 500
similar to those obtained using the continuous
Aluminum Reduction Cells’, Light Metals 1996, ed. kA Cell Design into a 600 kA Cell design’, ALU-
tracking feed control algorithm with shorter W. Hale (TMS, Warrendale, Pa), 451-457. MINIUM 87 (2011) 1-2, 52-55.
cycles, but with far less risk of having anode [2] M. Dupuis, I. Eick and F. Waldman, ‘Modelling
effects. Thermal Dynamic Response to a 3-Hour Total Power
Shutdown Event’, 9th Australasian Aluminium Smelt- Author
ing Technology Conference and Workshops, (2007).
Conclusions [3] M. Dupuis and M. C. Schneller, ‘Testing In Situ Dr. Marc Dupuis is a consultant specialised in the
Aluminium Cell Control with the Dyna/Marc Cell applications of mathematical modelling for the
The author hopes that this demonstration Simulator’, COM, (2011), to be published. aluminium industry since 1994, the year when he
study highlights the value of using a dynamic [4] Y. Macaudiere, ‘Recent Advances in Process founded his own consulting company GeniSim Inc
cell simulator to optimise existing cell control- Control of the Potline’, Light Metals, 1988, ed. L. (www.genisim.com). Before that, he graduated with
ler algorithms, and to test new ones, without G. Boxall (TMS, Warrendale, Pa.), 607-612. a Ph.D. in chemical engineering from Laval Univer-
[5] M. Dupuis and H. Côté, Dyna/Marc Version 13 sity in Quebec City in 1984, and then worked ten
putting real cells at risk. The Dyna/Marc cell User‘s Guide, (2011). years as a research engineer for Alcan International.
simulator used in this study is available to the [6] K. R. Robilliard and B. Rolofs, ‘A Demand Feed His main research interests are the development of
whole aluminium industry through GeniSim Strategy for Aluminium Electrolysis Cells’, Light mathematical models of the Hall-Héroult cell deal-
Inc. Version 13 included the linear and quad- Metals, TMS (1989), 269-273. ing with the thermo-electric, thermo-mechanic,
ratic RMS noise filtration algorithms and the in [7] M. C. Schneller, ‘In Situ alumina feed control’, electro-magnetic and hydrodynamic aspects of the
JOM, 61(2009)11, 26-29. problem. He was also involved in the design of ex-
situ feed controller algorithm. The Dyna/Marc [8] M. C. Schneller, ‘In Situ aluminum feed control’, perimental high amperage cells and the retrofit of
cell simulator can also be used as a cell design Light Metals 2010, ed. J. A. Johnson (TMS, War- many existing cell technologies. Contact: dupuis@
tool, as demonstrated in [10]. rendale, Pa), 563-568. genisim.com

ALUMINIUM · 1-2/2012 55