Work

Optimizing coal blending for metallurgical coke production using a reinforcement learning-enhanced evolutionary algorithm

Year: 2025

Type: article

Abstract: Achieving low-cost, high-quality coke production remains an ongoing challenge in the coking industry, where the coal blending optimization problem is recognized as NP-hard. Earlier approaches typically adopt single-objective optimization strategies, focusing on cost or selected quality indices but neglecting the need to balance both simultaneously. To address this challenge, we present a multi-objective coal blending optimization model (MOCB) that minimizes blending cost and maximizes critical coke quality parameters (ash, sulfur, M40, M10, CRI, CSR). We further propose a Q-learning enhanced Nondominated Sorting Genetic Algorithm-III (QNS- GAIII) to effectively explore the extensive solution space—on the order of 10020 when blending 20 types of coal—while maintaining the required process constraints. Benchmark tests on DTLZ and WFG functions show that QNSGAIII surpasses five classical champion algorithms (MOEA/D, PESA-II, RVEA, NSGA-II, NSGA-III) in terms of convergence and diversity (IGD, HV, etc.). An industrial case study involving three distinct metallurgical coke products demonstrates how the QNSGAIII-derived Pareto-optimal solutions simultaneously achieve cost savings and meet quality specifications, revealing robust and diverse coal blend structures. Optimization of coal blending reduces the use of premium coking and fat coals by 3%-5% and lowers blending costs by 10-20 RMB per ton. These results demonstrate the practical value of incorporating reinforcement learning into evolutionary algorithms for multi-objective coal blending, providing a route to more efficient and intelligent cokemaking processes.

Source: Fuel

DOI: 10.1016/j.fuel.2025.136777

Authors: Zhongfeng Li, Lei Liu, Zhenlong Zhao, Hui Jin, Yuting Zhuo

Institutions: University of Science and Technology Liaoning, Yingkou Institute of Technology, UNSW Sydney

Language: English

Topic: Advanced Multi-Objective Optimization Algorithms

Subfield: Computational Theory and Mathematics

Field: Computer Science

Domain: Physical Sciences

SDGs: Affordable and clean energy, Responsible consumption and production

Open Access status: hybrid

Estimated APC paid: $4,330