Adaptive physics-informed cascaded neural networks for nuclear reactor core parameter identification

Abstract

In nuclear power plant system simulations, simulation model accuracy directly influences dynamic characteristic analysis and control system design. To meet real-time requirements, system-level models typically employ simplified modeling approaches, whose internal structural parameters or physical property parameters deviate from the actual system, thereby reducing model accuracy. Furthermore, traditional parameter identification methods often struggle to effectively address time-varying parameters, particularly when operational data is sparse and power range coverage is incomplete. Therefore, this paper proposes an adaptive physics-informed cascaded neural network (PICNN) method for identifying physical property parameters that varies with reactor operating state. The validation results show that, under both noise-free and noisy data conditions, the proposed method has good parameter identification performance and robustness. Moreover, the model output optimized through parameter identification agrees well with the simulation model output data.