Smart dispatching for low-carbon mining fleet: A deep reinforcement learning approach

Abstract

As the mining sector strives to meet its greenhouse gas (GHG) reduction targets by 2030 and 2050, the need for efficient energy-saving measures becomes paramount. This study presents an improved smart dispatching solution developed using deep reinforcement learning to improve the performance of mining haul truck fleets. The proposed solution based on double deep Q-network (DDQN) trains each truck in the fleet to make real-time decisions based on various factors such as in-situ grade (ore or waste), road traffic, estimated queueing, and maintenance requirements. Scenarios based on a synthetic mine are simulated to evaluate the effectiveness of this DDQN-based artificial intelligence solution versus conventional approaches. Our results demonstrate that the proposed DDQN solution not only improves productivity and reduces fuel-related GHG emissions, but also significantly improves fleet performance in handling operational disruptions (production loss reduced by over 50%) without human intervention, including unexpected changes in fleet size and shovel grade. The cost-effectiveness and scalability of the proposed solution are also assessed. Comparative analysis revealed that upgrading the fleet with the proposed DDQN solution to reduce GHG emissions only equates, on average, 47% and 21% of fleet electrification and carbon capture and storage costs, respectively, and can be rapidly adopted in the near term. These advantages position the smart fleet dispatching system as a cost-effective approach to enhance productivity and reduce direct GHG emissions in mining operations for short-term targets, making it an adaptive technology for future automated and sustainable mining.