Evaluation of optimized machine learning models for nuclear reactor accident prediction

Several studies have proposed machine learning models to diagnose and predict accidents in nuclear power reactors. However, the training data in these studies are deterministic, and they don't consider the parameter uncertainty caused by sensor failure during an accident. The performance of the ML must be weighted with the uncertainties and possible malfunctions of the detectors to be considered feasible. Consequently, this work presents a novel training approach for machine learning models using an augmented dataset that reflects the sensor status. In this study, three machine learning models, Support Vector Machines (SVM), Decision Trees (DT) and Multilayer Perceptron (MLP), are developed, trained and compared. By simulating the loss of coolant accident (LOCA) and Steam Generator Tube Rupture (SGTR) accident, several augmented reactor coolant parameters were used to train and test the machine learning models to predict reactor accidents. The model performances were evaluated using the F1 score, precision, accuracy, learning curves, and the Confusion Matrix to determine the suitable algorithm for accurate diagnosis of the two reactor accidents. From this study, we conclude that the SVM and DT models performed better than MLP for these accident scenarios.