Simultaneous stochastic optimization of an open-pit mining complex with preconcentration using reinforcement learning
Zachary Levinson, Roussos Dimitrakopoulos, Julien Keutchayan
Abstract
A preconcentration facility is a major operational component that is critical for managing capacities and improving process plant efficiency in a mining complex. These facilities have not been considered in previous short-term production scheduling frameworks for mining complexes. Short-term production scheduling is a vital part of planning that helps ensure long-term production targets are meet without compromising value. In this work, a new stochastic mathematical programming formulation for simultaneously optimizing the short-term production schedule with preconcentration considerations is proposed. The optimization formulation considers optimizing the extraction sequence, destination policy, stockpiling and preconcentration decisions jointly to capture potential synergies. In addition, this work investigates a new approach for short-term production scheduling that combines reinforcement learning with stochastic mathematical programming. An actor–critic reinforcement learning agent learns to optimize the short-term production schedule and provides a more flexible framework for adapting heuristics to the scheduling problem. The optimization approach and stochastic formulation are tested in a copper mining complex with multiple mining areas, several material properties, stockpiles, preconcentration facilities, leach pads, process plants and waste dumps. The case study shows the practical aspects of the proposed optimization and the direct benefit of integrating preconcentration decisions in the short-term production schedule; this led to a $140M improvement in annual cashflow. Additionally, the actor–critic reinforcement learning algorithm learns a stable policy that provides operational extraction sequences.