component driven among other things by the fluctuation of the alumina concentration in
the bath, which is the relevant data for the alumina feeding control algorithm.
resistance [1]. In the current work, a cell voltage free of any amperage fluctuation noise
is recomputed using the nominal amperage. The cell voltage computed this way every 6
seconds is then averaged every 2 minutes. Finally, the best straight line fitting the last 10
"2 minutes averaged cell voltage" is computed. Every 2 minutes, a new 2 minutes cell
voltage averaged value is computed and a new best straight line fitting using the last 10
"2 minutes averaged cell voltage" is computed. Figure 2 presents the results obtained
after 40 of such cycles. The smoothed voltage curve presented in Figure 2 was
constructed using the red section of the fitted line presented in Figure 1 for each of those
40 cycles. Although it looks like a monotonous curve, Figure 3 presents the same
smoothed voltage curve highlighting the fact that there are jumps in the curves. Yet,
Figure 4 clearly shows that over a period of 12 hours, this algorithm filtered very well the
noise, generating a remarkably smooth low frequency voltage evolution curve. As
highlighted in Figure 3, the voltage slope evaluations are discrete, but as highlighted in
Figure 5, over a period of 12 hours, those discrete voltage slope evaluations are
producing a remarkably continuous curve very well suited to be used as input data for the
continuous tracking alumina control algorithm.
Figure 6 illustrates the second algorithm tested. The same amperage fluctuation
this time, the best parabolic curve fitting the last 10 "2 minutes averaged cell voltage" is
computed. Obviously, as it can be seen comparing Figure 6 to Figure 1, the global fit
over the those last 10 averaged points is better but does it means that this more complex
and more CPU demanding algorithm computes a more reliable voltage slope for the
continuous tracking alumina controller? Figures 7, 8 and 9 present the smoothed voltage
curve obtained. When compared with Figures 2, 3 and 4, we can see that taking the extra
trouble of computing the best parabolic curve fit instead of the best straight line fit did
not resulted in producing a smoother voltage curve, on the contrary! Figure 10 confirms
that the voltage slope computed with the second algorithm is a lot noisier. Furthermore,
the maximum amplitudes of the slope computed with the second algorithm are less
accurate than the ones computed with the first simpler algorithm.