Improving Accuracy of a Machine Learning Approach for Forecasting Severe Accidents

Abstract

Severe accidents require accurate and fast prediction tools to support real-time decision making. This study presents a machine learning-based framework for forecasting thermal-hydraulic (TH) variables during a loss-of-component-cooling-water (LOCCW) accident. Using datasets generated by the Modular Accident Analysis Program (MAAP), surrogate models were developed to predict key TH variables observable in the main control room. Two approaches, Multi-Input Multi-Output (MIMO) and Multi-Input Single-Output (MISO), were evaluated. The MISO approach outperformed MIMO, reducing regression errors by an average of 43% and improving full-scenario prediction accuracy in over 90% of test cases. This improvement is attributed to the MISO ability to eliminate gradient conflicts and specialize in individual variable predictions. However, challenges persist in accurately forecasting reactor vessel water level (RV WL) and maximum core exit temperature (MAX CET), which exhibit higher errors. Despite requiring more computational resources, MISO models significantly reduce computation time compared to traditional system codes, offering predictions within seconds. This study demonstrates the potential of MISO-based models for real-time accident forecasting and lays a foundation for further refinement and broader application to other accident scenarios.