Application of GMDH Method to Core Power and k_eff Prediction

This study develops surrogate models for a small reactor based on the Group Method of Data Handling (GMDH), a self–organizing neural network algorithm. A dataset has been constructed using the computational results from a high–fidelity, three–dimensional RMC model. A power prediction surrogate model has been trained, which takes axial three–section detector readings as input to predict 15-section power distributions. Concurrently, a k_eff prediction surrogate model has been established using 25 modeling parameters as input. The power prediction model achieves a relative deviation mostly within 3 %. By applying a piecewise polynomial fitting method, the resulting pellet–level power predictions show a relative deviation of less than 5 %. The k_eff prediction model demonstrates a root mean square error (RMSE) under 36 pcm, with the discrepancy between its predicted uncertainty and RMC's statistical results being less than 18 pcm. This work serves as a valuable reference for applying surrogate models to core power prediction and uncertainty analysis.