Fusing Neural Network Surrogates with EnKAPF for Real-Time State and Parameter Estimation in Nuclear Reactors

Real-time estimation of unmeasurable parameters, such as the heat transfer coefficient (HTC), is critical for accident diagnosis in nuclear reactors. This paper introduces an intelligent software sensor that addresses this challenge by fusing a neural network (NN) surrogate model with an advanced data assimilation algorithm—the Ensemble Kalman Auxiliary Particle Filter (EnKAPF). This framework enables simultaneous online estimation of the reactor core temperature and the unknown HTC. We rigorously validate the framework using experimental data from a Loss of Feedwater (LOFW) accident, assessing three key aspects of its performance. First, its tracking accuracy is demonstrated by a mean absolute percentage error (MAPE) of only 0.170% in temperature estimation. Second, the physical consistency of the estimated HTC is confirmed, as it drives the standalone NN model to reproduce experimental temperatures with a low MAPE of 0.514%. Finally, a low one-step-ahead prediction error validates its predictive power, essential for early-warning applications. Significantly, the framework successfully characterizes the dynamic HTC profile during a LOFW transient, clearly capturing the transition from forced to natural circulation. By enabling such deep diagnostics, the proposed sensor provides core technology for advanced fault detection, precise accident analysis, and the development of high-fidelity digital twins for nuclear reactors.