Operation Optimization Framework for Advanced Reactors Using a Data-Driven Digital Twin

Abstract

Abstract To meet future energy demand, while producing safe, reliable, and carbon-free energy, nuclear reactors will be needed. To make next-generation reactors more economically competitive, the Artificial Intelligence and Machine Learning-based operation optimization module of the dynamic operation and maintenance optimization (DyOMO) framework is proposed. The operation optimization module of DyOMO consists of a data-driven digital twin coupled to a genetic algorithm (GA) optimizer to quickly and efficiently search the solution space for optimal control schemes. The digital twin consists of a Bayesian Network (BN) known as MVCBayes to incorporate uncertainty in the optimization and feedforward neural networks (FFNN) as MVCNet with GA to conduct the optimization. Over time as reactor systems are operating, component degradation will cause the system's electrical output to decrease or be shutdown entirely to perform maintenance. To prevent this, the DyOMO operation optimization module aims to prolong system operation until the next scheduled maintenance period using multiple variable control (MVC) that simultaneously perturbs all actuators to better control the reactor. Comparing this approach with traditional single variable control (SVC), MVC can extend reactor operation past 5% degradation while SVC begins to struggle once the pump and turbine degradation surpasses 0.85% for load-following (LF) operation. Given this extra operation time, the system can continue to run while maximizing its safety margin until the next scheduled shutdown and potentially decrease the total number of maintenance actions throughout the license period to decrease operational and maintenance (O&M) costs.