ENCYCLOPÉDIE DE LA RECHERCHE SUR L’ALUMINIUM AU QUÉBEC 2013 | Page 33

PRODUCTION DE L’ALUMINIUM

Application of an Artificial Neural
ALUMINIUM PRODUCTION
Network to Control the Properties of
APPLICATION D’UN RÉSEAU DE NEURONES ARTIFICIELS
Baked Anodes

31

POUR CONTRÔLER LES PROPRIÉTÉS DES ANODES CUITES

(Application d’un réseau de neurones
APPLICATION des
artificiels pour contrôler les propriétésOF AN ARTIFICIAL NEURAL NETWORK
TO CONTROL THE PROPERTIES OF BAKED ANODES
anodes cuites)

Dipankar Bhattacharyay1, Duygu Kocaefe1, Yasar Kocaefe1, Brigitte Morais2
Auteur 1, Author2
1Université

2Aluminerie

du Québec à Chicoutimi, Département des sciences appliquées, 555, boul. de l’Université, Chicoutimi, Québec, Canada G7H 2B1
1 Département et Institution 1
Alouette Inc., 400, Chemin de la Pointe-Noire, C.P. 1650, Sept-Îles, Québec, Canada, G4R 5M9
2 department and Institution 2.

CHAIRE DE RECHERCHE UQAC/AAI SUR LE CARBONE

Artificial Neural Network

Flow Sheet
Experimental Data
Normalization of Data

Abstract
The carbon anodes contribute significantly to the cost of primary
aluminum production. The variations in quality of raw materials such as
calcined coke, pitch, recycled butts and anodes affect the quality of
baked anodes. The goal of the industry is to produce better quality
anodes in spite of these variations. In the absence of a distinct
mathematical relationship between the raw material properties and those
of the baked anodes, the production is usually controlled based on
experience. The plants usually maintain a large database. Proper
analysis of these data using artificial neural networks (ANN) can help
better control the anode production process and improve the anode
quality. In this work, the application of an ANN to control the quality baked
anodes is demonstrated in the case of variations in raw materials using
data from the literature.

Results and Discussion

Develop Artificial Neural Network Model

Initialization
Training of the network
Adjustments of weight w and bias b

The results indicate that it is possible to replace Coke D with other
cokes without affecting significantly the baked apparent density,
specific electrical resistivity, and Young’s modulus of the baked
anodes.

Learning
Pattern recognition and data classification
Validation
Compare output with test data set
Prediction
Predict output using data set for which output is not known

Data Used for Analysis

Data from a published doctoral thesis (Juraj Chmelar, Size
reduction and specification of granular petrol coke with respect to
chemical and physical properties, The Norweigian University of
Science and Technology, 1992) were used for the analysis.
Chemelar used 4 cokes mixed with pitch (15 to 20%), varied the
percentage of -63µm coke particles (45 to 94%), and measured
different baked anode properties.

Properties of Cokes Used

Predicted vs. Experimental Values

Conclusions

ANN is an useful tool for the prediction of anode properties.
It can be used as a tool to control the quality of anodes when there are
variations in raw materials and process conditions.
The major advantage of ANN is that it can efficiently handle highly
nonlinear data with noises for which there is no existing mathematical
relationship.
The development of an efficient ANN model is time consuming because it
needs lot of trials and errors; but, once it is developed, it can predict results
even if there are no experimental data available.
For the training of an ANN, availability of large sets of data is required, but
this is generally not a limitation for the industries.

Dipankar Bhattacharyay
Duygu Kocaefe
Yasar Kocaefe
Université du Québec
à Chicoutimi
Brigitte Morais
Aluminerie Alouette Inc.

Acknowledgements

The technical and financial support of Aluminerie Alouette Inc. as well as the financial support of the
National Science and Engineering Research Counc