Two-stage distributionally robust optimization for joint system design and maintenance scheduling in high-consequence systems

• Published in: IIE transactions Vol. 56; no. 8; pp. 793 - 810
• Main Authors: Zhang, Hanxiao, Li, Yan-Fu, Xie, Min, Zhang, Chen
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis Ltd 02.08.2024
• Subjects: Adaptive sampling; Degradation; Design optimization; Feedwater; Maintenance management; Nuclear power plants; Nuclear safety; Optimization; Preventive maintenance; Risk aversion; Robust design; Scheduling; Search algorithms; Stochastic programming; System reliability; Systems design
• ISSN: 2472-5854; 2472-5862
• DOI: 10.1080/24725854.2023.2225097

The failures of high-consequence systems can cause serious harm to humans, including loss of human health, life security, finance, and even social chaos. To protect high-consequence systems, both optimal system design and maintenance activities contribute to improving system reliability and social safety. The existing works generally optimize these two problems sequentially and assume that the degradation process of components is precisely known. However, sequential optimization often results in significant losses due to redundancies, and such a presumption usually cannot be guaranteed in practice, due to limited historical data or a lack of expert knowledge, referred to as epistemic uncertainty. To fill this gap, in this article, we consider an integrated optimization of system design and maintenance scheduling for multi-state high-consequence systems in which the component’s degradation is known with limited distributional information. To address this issue, we utilize the framework of distributionally robust optimization to provide a risk-averse decision to decision-makers even under the worst realizations of random parameters, and develop a two-stage integer distributionally robust model with moment-based ambiguity set to determine the system design and maintenance scheduling simultaneously. The proposed model can be converted to a tractable approximation as an integer linear stochastic programming problem. In order to solve large-scale problems, we develop a sample-based adaptive large neighborhood search algorithm to find the optimal system designs. In the numerical experiments, we present a case study on feedwater heating systems in nuclear power plants and demonstrate that an integrated optimization consideration creates significant benefits in profitability. We also present the out-of-sample performance of the distributionally robust design to avoid extreme risk.