No abstract

In this work, neural networks are built and trained to be used in a predictive control scheme for the aluminum electrolytic cell. An efficient control of the cell requires the knowledge of predicted future values of the decision variables in order to enable the standard (nonpredictive) control logic to take anticipated actions to prevent the anode effect, a destabilizing event occurring during cell operation. The networks are first trained on data obtained from a computer simulator of the cell prior to undergoing further on-line learning. Trained to predict the cell resistance and the resistance's trend indicators, the networks are applied to the control of cells in different cell states, with a view to preventing anode effects, the latter being deliberately induced by reducing the alumina feed rate or reducing the feeding frequency and duration. Results show that, with neural-predictive control, anode effects can be avoided, which should result in increased thermal stability, decreased power consumption, and reduced fluoride emissions. Further, the on-line learning capacity of the networks offers a good perspective for their application to other complex industrial processes as well.

No abstract

To be efficient, the control of alumina feeding of the electrolytic cell must be based on cell resistance, alumina concentration, and cell state. Most control schemes now in use are based on cell resistance only, and, thus, constitute an open-loop control that lacks robustness because their decision criteria are not explicitly tied to concentration nor to cell state. This results in the cell operating at nonoptimal concentrations, and cell efficiency is diminished. An optimal operation requires a knowledge of concentration and an adjustment of the decision criteria as a function of concentration. A learning vector quantization (LVQ) type of neural network was built and trained to recognize the cell state. Knowing the state of the cell and its resistance, concentration can be estimated using predetermined regression functions. The decision criteria for the control logic are then consequently adapted. A closed-loop control scheme is thus obtained. Results show that, with its control so structured, the cell can operate at or near optimal concentrations independently of its state. This flexible and intelligent character of the neural control can provide a considerable advantage as compared to the standard control.