Multi-objective genetic algorithm embedded with reinforcement learning for petrochemical melt-flow-index production scheduling

Abstract

In the petrochemical industry, product type conversion can cause capacity loss and produce transition products, leading to waste. Our study aims to minimize these transition products and tardiness in meeting customer due dates. We introduce the Engineering Experience Heuristic (EEH), which utilizes scheduling reference sheets to identify key production traits and constraints, thereby generating efficient schedules. Further building on the EEH, we propose a non-dominated sorting genetic algorithm II enhanced with reinforcement learning (NSGAeRL) to address the multi-objective scheduling optimization problem. The incorporation of reinforcement learning (RL), particularly through a model-based Markov Decision Process (MDP), allows for the dynamic adjustment of genetic algorithm (GA) parameters, including crossover and mutation mechanisms, based on the optimal policy derived from RL. We validate our proposed methodologies through empirical studies and numerical experiments. The results demonstrate that NSGAeRL outperforms standard benchmarking algorithms, offering superior performance in minimizing transition products and scheduling tardiness.