Real-time Grinding Energy Consumption Forecast and Control with Machine Learning

Abstract

Comminution is the most energy-intensive stage in mineral processing and accounts for up to 50% of the total energy consumption in mining operations. This stage, particularly the grinding process, involves a series of strategically arranged mills—rod and ball mills in this case—designed to reduce the material. In an open circuit rod mill and closed circuit ball mill configuration adjusting specific key mill parameters can significantly influence the energy consumed. This research presents an approach to forecasting power consumption in rod and ball mills using machine learning models. The study focuses on predicting power usage 20 minutes into the future and interpreting the impact of various mill parameters on power efficiency. The machine learning models developed forecasted the power consumption, achieving R2 values of 0.82, 0.73, 0.84, and 0.86 for XGBoost, Random Forest Regressor, Support Vector Regressor, and Deep Neural Networks models, respectively, for the rod mill. For the entire circuit, including both rod and ball mills, the R2 scores were 0.74, 0.73, 0.70, and 0.50 for the respective models. The study highlights the critical role of optimizing grinding circuits to enhance energy efficiency in mineral processing. Given the high cost of the milling process, inefficiencies in grinding circuits, particularly those using rod and ball mills, result in significant energy waste. Unlike previous studies, which primarily focus on predicting power consumption in SAG mills and optimizing production without considering energy constraints, this research presents a machine learning-based recommendation system to achieve power reduction while maintaining production level and particle size targets. Utilizing historical data from a mineral processing plant, a reinforcement learning-based recommendation engine was developed to suggest adjustments to the rod mill's water flow, achieving a projected power reduction of up to 2%, equivalent to 299.24 Megawatt hours annually. This research emphasizes the considerable potential for enhancing sustainability and reducing costs in mineral processing through the utilization of machine learning to adjust mill parameters.