A critical component of the autonomous control system is the implementation of digital twin (DT) for diagnosing the conditions and prognosing the future transients of physical components or systems. The objective is to achieve an accurate understanding and prediction of future behaviors of the physical components or systems and to guide operating decisions by an operator or an autonomous control system. With specific requirements in the functional, interface, modeling, and accuracy, DTs are developed based on operational and simulation databases. As one of the modeling methods, data-driven methods have been used for implementing DTs since they have more adaptive forms and are able to capture interdependencies that can be overlooked in model-based DTs.
To demonstrate the capabilities of DTs, a case study is designed for the control of the EBR-II sodium-cooled fast reactor during a single loss of flow accident, where either a complete or a partial loss of flow in one of the two primary sodium pumps is considered. Based on the definition of DTs and the design of autonomous control system, DTs for diagnosis and prognosis are implemented by training feedforward neural networks with suggested inputs, training parameters, and knowledge base. Furthermore, inspired by the validation and uncertainty quantification scheme for scientific computing, a list of sources of uncertainty in input variables, training parameters, and knowledge base is formulated. The objective is to assess qualitative impacts of different sources of uncertainty on the DT errors. It is found that the performance of DT for diagnosis and prognosis satisfies the acceptance criteria within the training databases. Meanwhile, the accuracy of DTs for diagnosis and prognosis is highly affected by multiple sources of uncertainty.