Abstract

Comminution is the most energy-intensive phase in mineral processing, with up to 50% of a mining complex's total energy dedicated to this stage. Controlling the energy consumption in circuits that contain rod and ball mills is a significant challenge. Fine particle overgrinding is one of the biggest causes of energy waste as well as creating further issues downstream, influenced by factors such as water levels affecting the slurry density and cyclone efficiency. Careful control of equipment parameters is required to optimize energy usage in grinding circuits.

A new recommendation system approach was developed using machine learning to monitor and reduce the energy consumption in grinding circuits while maintaining the production throughput. The relationship between energy and the grinding circuit parameters such as particle size and water addition is not very well understood; therefore, supervised machine learning models including random forest and XGBoost were used to monitor key equipment parameters and observe their impact on energy consumption. The models are trained on historical data to mimic the system's functionality, allowing to use them as surrogate models. By fine-tuning the algorithms, energy consumption prediction accuracy improved by up to 73%, allowing a better understanding of grinding circuit equipment behaviour and the factors affecting energy consumption.

Building upon this, an additional machine learning approach, deep Q learning which is a type of reinforcement learning (RL) was implemented to reduce energy consumption while maintaining the desired throughput levels and target grind size. After training on historical data, the RL model makes sequential decisions in a complex grinding circuit environment by proposing actions including increasing, keeping or decreasing the water addition to the rod mill. By applying rewards and penalties to the decisions, the trained reinforcement learning model aims to find the optimal sequence of actions to reduce the energy intensity and not compromise the production quality. The setup of the RL action and state spaces is determined using the observations made on the key mill parameters previously identified by the surrogate model.

Our case study has shown that by utilizing supervised models to understand equipment behaviour and as well as integrating RL, the recommendation system was potentially able to reduce the energy consumption in grinding circuits by up to 2%, as calculated using the surrogate model,  benefiting from machine learning's capacity to learn and adapt to complex operational environments.