The accuracy of transient thermal hydraulic parameter prediction of reactor cores under various working conditions directly affects reactor safety. Mass flow rate and temperature are important parameters of core thermal hydraulics, which are often modeled as time-series prediction problems.

This study aims to solve the accuracy problem of continuous prediction of core thermal hydraulic parameters under instantaneous conditions and to test the feasibility of a gated cycle unit based on the attention mechanism in core parameter prediction.

The 1/2 full core model of China Experimental Fast Reactor (CEFR) core was taken as the research object. The subchannel SUBCHANFLOW program was employed to generate the time series of transient core thermal hydraulic parameters. The gated recurrent unit (GRU) model based on soft attention was used to predict the mass flow and temperature time series of the core.

The results show that, compared with the adaptive radial basis function (RBF) neural network, the GRU network model with soft attention offers better prediction results. The average relative error of temperature is <0.5% when the step size is 3, and the prediction effect is quite good within 15 s. The average relative error of mass flow rate is <5% when the step size is 10, and fairly good prediction effect is achieved in the subsequent 12 s.

The model constructed in this study not only exhibits higher prediction accuracy in the continuous prediction process but also captures the trend characteristics in the dynamic time series, which is of considerable value for maintaining reactor safety and effectively preventing nuclear power plant accidents. The GRU model based on soft attention can provide continuous prediction for a period of time under transient reactor conditions, providing a reference value in engineering applications and improving reactor safety.