Comparative Machine Learning Study for Estimating Peak Cladding Temperature in AP1000 Under LOFW

For a more realistic estimation of safety margins, the conservative approach is replaced by integrating the best estimate approach (BE) with uncertainty quantification, the integration which knows as best estimate plus uncertainty (BEPU), which can predict the key safety parameters such as peak cladding temperature (PCT) and departure from nucleate boiling ratio (DNBR), etc. In this sense, a fast and cost-effective tool for uncertainty quantification is developed through a data-driven approach to predict PCT under loss of feedwater accident (LOFW) in AP1000 reactor. This paper includes performing a comparative study between different regression ML algorithms to find the best algorithm which can predict the PCT with higher accuracy. Intent to generate the required data for training and testing the ML algorithm, an uncertainty quantification framework is developed by coupling a best estimate code (RELAP5) with a statistical tool (RAVEN). RELAP5 is used to simulate the thermal-hydraulic response under LOFW accident while a set of uncertainty parameters are propagated through the RELAP5 model using RAVEN. These distributions were sampled using a Latin Hypercube Sampling (LHS) technique to generate sets of sample cases to simulate using the RELAP5 code. 5,000 runs were generated in order to acquire a large database for training purposes. The examined algorithms are linear regression, supported vector machine, k-nearest neighbors (KNN), and random forest. The evaluation of algorithms depends mainly on mean absolute error (MAE) and determination coefficient R2. The result shows that the random forest provides high accuracy in predicting PCT within four algorithms, which reaches 98.96%.