amperage and voltage and then use that information to send instructions to the point breaker feeder
and the anode beam in order to keep both the dissolved alumina concentration in the bath and the
anode cathode distance (ACD) under tight control.
simulated cell instead of conducting those tests on real cells. This is true as long as the behavior of
the simulated cell is reliable enough to provide useful feedback. In order to achieve that goal, the
Dyna/Marc cell simulator has been continuously improved since 1994. It already demonstrated its
ability to reproduce measured cell dynamic evolution in previous publications [1, 2].
been offering the option to add amperage and voltage noise to the simulation. For the cell voltage
that is an output to the simulation, the noise generated by the bath-metal interface motion and the
bubble release is added to the calculated noise free voltage at the end of each time step. The level of
the added noise is function of the ACD, the thickness of the metal pad, the amount of sludge and the
fraction of the anode surface covered by frozen bath. This noise level, that can be made to affect
current efficiency, can be reduced by automated voltage treatment since version 1.4 issued in 1999.
resistance as it would lead to useless results. Since version 13.0 issued in 2011, Dyna/Marc is
offering linear and quadratic root mean square (RMS) noise filtration algorithms [3]. Figure 1 is
showing the comparison between the noise-free and the noisy evolution of the cell pseudo-
resistance. The aim of the cell controller noise filtration algorithm is to use the noisy data to
estimate the evolution of the slope of the noise-free curve. Figure 2 is showing the comparison
between the noise-free slope evolution and the slope evolution estimated using three different
modes of filtration.
themselves 5 seconds averaged value of the raw cell voltage measured at a 10 Hz frequency. As it
can be seen, the resulting estimation is still a bit noisy. The second one in the middle was obtained
using 120 datapoints instead of 60 datapoints. The result is almost noise-free but now the estimation
is dragging 5 minutes behind the noise-free 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 represents the state of the slope 5 minutes ago. In the example presented in Figure 2, the slope
is doubling in 5 minutes during a no-feed observation, so the estimated value is noise-free but about
half of the real value. The third mode of filtration on the right of Figure 2 was obtained using
quadratic RMS fitting also using 120 datapoints. Using quadratic RMS fitting of the cell voltage
evolution eliminates the drag in the slope estimation which is important, but for the same number of
datapoints used, generates a more noisy estimation.
feed control algorithms are based on continuous tracking or underfeeding and overfeeding cycles
where the shift from underfeeding to overfeeding is dictated by a trigger value or either the slope of
the cell pseudo-resistance or the slope of the cell normalized voltage. One of the earliest versions of
that algorithm can be found in Figure 3 of Aluminium Pechiney 1988 TMS paper [4]. That
algorithm is available in Dyna/Marc simulator under the name Pechiney Tracking Feed Control [5].