Dynamic identification and rapid power prediction of a microreactor helium-xenon Brayton cycle under reactivity perturbations

Objectives This study aims to enable rapid electrical power prediction for a gas-cooled microreactor helium–xenon closed Brayton cycle under external reactivity perturbations, thereby reducing the computational cost of long-duration dynamic analyses across multiple operating conditions. Methods A dynamic system model was used to generate time series under random reactivity perturbations, with external reactivity as the input and electrical power as the output. Ridge regression and autoregressive with exogenous input (ARX) models were constructed alongside nonlinear autoregressive with exogenous inputs (NARX) models. The nonlinear mappings were implemented using a backpropagation (BP) neural network and a temporal convolutional network (TCN), yielding NARX-BP and NARX-TCN, respectively. Independent random test sequences were used to compare prediction accuracy and error distributions within the training range and under an extended reactivity amplitude range. The computational efficiency of NARX-TCN was also evaluated against the dynamic system model. Results The nonlinear NARX-BP and NARX-TCN models generally outperformed the linear ridge regression and ARX models in prediction accuracy and adaptability to the extended amplitude range. When the reactivity amplitude range was extended from ±0.15 <inline-formula> to ±0.25 </inline-formula>, the coefficient of determination of NARX-TCN decreased from 0.9964 to 0.9909, a reduction of only 0.0055. Under the extended range, its root mean square error was 6.30W, the lowest among the four models, and its error distribution remained relatively concentrated. In terms of computational efficiency, NARX-TCN achieved a speedup of approximately 1.97 × 10³ relative to the dynamic system model.Conclusions NARX-TCN substantially reduces computational cost while maintaining high electrical power prediction accuracy and provides the best adaptability among the four models under the tested amplitude extension. It can support dynamic analysis and rapid electrical power prediction for the underwater propulsion power system.