Additonal authors: Ortiz, J. M.. Book title: Proceedings of the 58th Conference of Metallurgists Hosting Copper 2019. Chapter: . Chapter title:
Semi-autogenous grinding mills are intensive energy consumers. Current models base their energy consumption inference on operational variables and feeding ore characterization. While they provide adequate design guidelines, they are not suitable for real-time energy forecasting dealing with gaps in operational dataset and up/downstream process bottlenecks. This work explores the capability of Gated Recurrent Units to learn the semi-autogenous grinding mill consumption behaviour. Gated Recurrent Units are trained to forecast different time-support consumptions under a time-series operational dataset. The results show high performance on small supports (30 minutes) to capture local variability and long-trends while larger time supports (8 hours) can only capture time-trends but not local variability. The proposed technique seems promising for further research and implementation on real mine cases since the input variables are low-cost and easy to obtain.
INTRODUCTION
Current changes in the Chilean energy matrix (Román-Collado, Ordoñez, & Mundaca, 2018), from fossil fuels to renewable energies, have impacted different industries and particularly the mining sector. Solar energy incorporation into greenfield/brownfield projects will help reducing costs and will lead to rethink foundational paradigms (production and processing strategies) (Pamparana et al., 2017). Also, integration of new space-time predicting tools together with automated real-time updating models are some of the main challenges in the geometallurgical framework (Ortiz et al., 2015; van den Boogaart & Tolosana-Delgado, 2018). Those contexts encourage and demand the development of tools capable of predicting the energy consumed by mining systems, being comminution the greatest energy consumer with an average close to
50% of the entire mine consumption (Cochilco, 2013).
In comminution, the semi-autogenous grinding mill (SAG) represents the largest energy consumer
(~10 𝑀𝑊 nominal each SAG). Theoretical and empirical energy consumption models (Jnr & Morrell, 1995; Morrell, 2004; Silva & Casali, 2015) base their inferences mainly on feed/product size distributions, SAG sizing, bearing pressure, feed hardness, water addition and grinding charge level, and commonly assume steady-state and work isolated from up and downstream processes. While those techniques provide adequate design guidelines, they are not suitable to predict hourly, daily or even weekly energy consumptions on interconnected comminution circuits where upstream/downstream bottlenecks lead SAGs to operate below designed regimes. Forecasting energy consumption is required by short-term energy management, medium and long-term planning energy consumption, designing the nominal capacity of renewable energy sources, among other. Also, when models are input-sensitive and some of those input measurements are expensive and/or time consuming, model robustness suffers when real-time predictions are required.
