Document Type : Original Research Paper

Authors

• Kusnawi Kusnawi ¹
• Mochamad Agung Wibowo ²
• Ridwan Sanjaya ³

¹ Doctoral Program of Information Systems, Universitas Diponegoro, Semarang, Indonesia ; Informatics, Universitas Amikom Yogyakarta, Yogyakarta, Indonesia

² Postgraduate School, Universitas Diponegoro, Semarang, Indonesia

³ Department of Information System, Universitas Katolik Soegijapranata, Semarang, Indonesia

Abstract

Open-pit haulage must reduce fuel use without sacrificing material movement or operational safety. This study examined whether a telemetry-driven, task-specific Digital Twin surrogate could support vehicle-level eco-driving policies that improve fuel intensity while preserving haul-cycle productivity. The framework combined full haul-cycle reconstruction, an ensemble XGBoost fuel-rate surrogate, nonlinear state-transition models, equal-budget multi-seed Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC), reward-sensitivity screening, a rule-based benchmark, and uncertainty- and out-of-distribution-aware evaluation. One-day telemetry from 71 Komatsu HD785-7 trucks contained 1,927,867 records and yielded 2,386 validated haul cycles. The fuel surrogate achieved a temporal-test MAE of 27.33 L/h, RMSE of 37.75 L/h, and R² of 0.6608. Across 26 complete independent test cycles, common-estimator rescoring showed that PPO reduced fuel intensity by 8.79% relative to the surrogate-human baseline (0.2854 versus 0.3163 L/t; p=0.0029), with a 4.91% mean throughput decrease that was not statistically significant. SAC reduced fuel intensity by 9.53% (0.2820 L/t; p=0.0032) and increased mean throughput by 3.16%, although the throughput change was not significant and loaded-route safety-envelope exceedance increased significantly. PPO and SAC did not differ significantly in fuel intensity. Policy-generated states also showed substantial out-of-distribution exposure. The framework therefore identified efficiency–productivity–safety trade-offs rather than universal algorithm superiority and should be interpreted as an offline policy-screening tool pending operator-in-the-loop and controlled field validation.

Keywords

• Digital Twin Surrogate
• Deep Reinforcement Learning
• Fuel Intensity
• Haul-Cycle Productivity
• Mining Haul Truck

Main Subjects

• Simulation & Optimization