Improving Accuracy of a Machine Learning Approach for Forecasting Severe Accidents

Abstract

As the progression of a severe accident is highly nonlinear and complex, there is a need to develop an accurate and accelerated prediction method that can support the decision-making of the operators. With this background, a machine learning-based severe accident prediction tool was designed. The main objective of this study is to find a time series forecasting strategy that enables an accurate prediction of thermal-hydraulic (TH) variables in a loss-of-component-cooling-water (LOCCW) accident. The datasets used for training the surrogate model were produced by the Modular Accident Analysis Program (MAAP). The surrogate model receives a set of information containing: ten TH variables that can be observed in the main control room, whether the safety components have failed or not, and whether the mitigation actions are implemented or not. Depending on whether the model predicts one TH variable or all ten TH variables, the model can be classified as a Multi-Input-Single-Output (MISO) or Multi-Input-Multi-Output (MIMO) model. Three types of deep neural network architectures have been tested to construct the surrogate model. It was found that the prediction accuracy of the ten TH variables was enhanced in most of the test scenarios by adopting the MISO strategy instead of the MIMO strategy. Not only were the peak values of the time series predicted more accurately, but the divergence issue at the latter part of the scenario was also resolved.