Assessing Machine Learning and Deep Learning-based approaches for SAG mill Energy consumption

Energy consumption represents a high operational cost in mining operations. Ore size reduction stage is the main consumer in that process, where the Semiautogenous Mill (SAG) is one of the main components. The implementation of control and automation strategies that can achieve production goals along with energy efficiency are a common goal in concentrator plants; however, designing such controls requires a proper understanding of process dynamics which are highly complex. This work studies machine learning and deep learning strategies that can be used to generate models for predicting energy consumption, using key process variables. In particular, the application of K-Nearest Neighbors Regressor (KNN-reg), Polynomial Regression (PR), Support Vector Regression (SVR) and Long-Short Term Memory (LSTM) strategies for energy prediction over SAG mill process data is developed in order to identify configurations suitable to be implemented for real-time prediction integrated over industrial data infrastructures. All techniques are compared in terms of Root Mean Square Error (RMSE) where, although all the models achieved acceptable performance, best results were obtained by a LSTM implementation which yielded an error of less than 4% associated to energy prediction.