Simulation-based exploration of the model of a Molten Salt Fast Reactor for the identification and classification of abnormal operating conditions

The deployment of the Gen-IV Molten Salt Fast Reactor (MSFRs) requires the demonstration of its enhanced safety features with respect to other reactor concepts. To this aim, a power plant simulator including the primary system, the secondary system and the balance-of-plant has been developed in the framework of the SAMOFAR EU project. This simulator allows to describe the plant (transient) response to a variety of (normal and abnormal) operating conditions. Within this framework, the objectives of the present thesis are: 1) to propose a simulation-based method to characterize the system behaviour with respect to variations in physical and operational parameters, by means of a thorough exploration of the MSFR power plant state space; 2) to develop a data-driven algorithm for the efficient detection and classification of incidents, relying on a k-Nearest Neighbors (kNN) classifier. The proposed approach comprises the following steps. First, a set of physical (input) parameters that are found to strongly influence the behaviour of the plant simulator (e.g., the fuel and intermediate salt mass flow rates and the gas flow rate) is selected, together with their ranges of variation. Second, several possible combinations of physical parameters values are generated by random sampling and the corresponding time-varying (transient) behaviour of the MSFR is simulated. Finally, the time evolution of some relevant (output) plant parameters (e.g., the fuel and intermediate molten salt temperatures) is analysed in detail to: (i) identify normal and abnormal system (output) configurations; (ii) train, validate and test the kNN incident detection and classification model; (iii) retrieve those combinations of the reactor physical (input) variables (e.g., circulation pumps failures) that are responsible for the abnormal system states (namely, fault diagnosis). The proposed method has shown a satisfactory performance: in particular, the incident detection and classification accuracy ranges between 89% and 99%.