Hybrid modeling method for reactor coolant loop combining data-driven and physics-based constraints

After a reactor accident, the complexity of the system and the large number of coupled system states hinder rapid and reliable accident diagnosis. Reactor modeling can assist operators by predicting accident evolution based on limited early-stage monitoring data. However, most existing data-driven models suffer from physical inconsistency, high uncertainty, and limited interpretability, while fully physics-based models are computationally expensive and difficult to deploy for early-stage accident prediction. To address these challenges, this study proposes a physics-constrained hybrid modeling framework, referred to as Hybrid Physical-Data Modeling (HPDM), for system-level reactor accident system states prediction. In the HPDM framework, mass, momentum, and energy conservation equations are incorporated into neural network training as physical constraints, allowing the model to simultaneously approximate observational data and fundamental physical laws. The HPDM method is applied to predict system states—time-dependent variables such as density, velocity, and internal energy—of a simplified reactor coolant loop. RELAP5 simulation results are used as the ground-truth reference, serving as a benchmark for quantitatively evaluating HPDM predictions against conventional purely data-driven models. The results show that HPDM significantly reduces the mean relative error, particularly in predicting the heater inlet flow rate, where the MSE decreased by 7 % compared to the conventional PINN method. The prediction uncertainty is also reduced, demonstrating its reliability and superiority in forecasting reactor accident progression.