Neural networks for the identification of the aluminium electrolysis process

Abstract

The proper operation of an aluminium electrolysis cell relies strongly on the ability of the control system to devise an appropriate alumina feeding sequence and on the ability to overcome anodic effects. This complex system is normally controlled using a rule-based control algorithm that stems from the process knowledge acquired over many years of operation. This paper first considers the use of feedforward neural networks to predict, 15 min in advance, two important variables that are directly used in the rule-based control algorithm: the cell resistance and the cell fast dynamic trend indicator. The addition of these two predictions greatly enhanced the operability of the cell. The linear vector quantization (LVQ) neural network was then used to classify the current operating state of the electrolysis cell in view of adapting the rule-based control strategy to each particular state. This classification procedure allowed to devise more appropriate control strategies and to estimate the alumina concentration of the cell.