Optimization of Aluminum Melting Process by Machine Learning Models for Time Series Data

Abstract

  Improving the melting efficiency in aluminum melting furnaces has significant cost and environmental benefits. A large amount of data, including time-series data, have been accumulated from the use of aluminum melting furnaces, but no effective method has been established to utilize these data. In this study, a data-driven model was constructed by combining two machine learning methods : variational autoencoder (VAE) and artificial neural network (ANN). VAE was applied as a model to quantify time series data into 18 latent variables, while ANN was constructed as a model to predict fuel gas consumption from latent variables and other characteristics. In addition, we attempted to optimize aluminum melting process by simulation using the data-driven model.

  Although the aluminum melting process was complicated, we were able to construct a highly accurate prediction model (R² ＝ 0.69). Furthermore, the characteristics of the fuel gas flow rate in the case of high melting efficiency were determined by simulation. In fact, the results of modifying the operating conditions of melting furnace based on the knowledge obtained confirmed a significant improvement in melting efficiency. These results indicate that the data analysis method used in this study is effective for process optimization.