A surrogate-assisted evolutionary algorithm for robust scheduling of nuclear power plant construction under project risk interactions

Nuclear power plant (NPP) projects are prone to major schedule overruns, often triggered by cascading risks such as supplier delays, human errors and change orders. Conventional scheduling models that treat risks as independent, fail to capture the systemic interactions that amplify disruptions to the project. This work introduces a surrogate-assisted evolutionary algorithm for robust project scheduling under uncertain risk interactions. A probabilistic Risk Interaction Network (RIN), calibrated on empirical evidence from European Pressurized Reactor (EPR) projects, is coupled with an Independent Cascade model and Monte Carlo simulation to assess the possible (stochastic) project outcomes. To address the issue of high computational cost, a neural network model ensemble is trained to approximate expected project durations from schedule–RIN configurations, whose predictive variance is used as an uncertainty constraint during project schedule optimization. Application to a Double Containment Pressurized Water Reactor (DC-PWR) case study shows that surrogate-guided search uncovers schedules with significantly lower expected duration and tail risks than deterministic baselines do, providing a tractable and risk-aware framework for robust NPP construction scheduling. • Formulates the robust RCPSP under risk interactions with a probabilistic RIN calibrated on EPR. • Couples RIN with Independent Cascade and Monte Carlo to quantify cascades leading to project delays. • SAEA is used for schedule optimization based on an ensemble neural network surrogate model with variance constraint. • DC-PWR case: robust schedules cut expected project makespan by >30 % with modest trade-offs.