Predication of 400kA Aluminium Electrolytic Cell Output Based on the Generative Adversarial Networks

It is difficult to apply accurately in the aluminum reduction industry that aluminum output is mainly based on the material balance and the experienced adjustment according to the performance of electrolytic cell. Dataset is established with data from 10 electrolytic cell in an electrolytic aluminum enterprise during 800 days. The generative adversarial network model inputting a 16-day serial data including the quality of molten aluminum, the condition of electrolytic cell, and the quantity of material, is built to predicate and evaluate aluminum output. In the experiment the models based on GAN to predicate aluminum output are trained by single and multiple datasets, and compared with BP and SVM models in performance. The results indicate that the model trained with combined dataset perform better than the single and the model based on BP and SVM, with the prediction accuracy of 99.4% and the maximum absolute error of 49.7.