A Control-Oriented Hybrid Model for Nuclear Reactors Based on Neural Network

Abstract Accurate modeling is the basis for analyzing the dynamic response characteristics of the model. However, due to the complexity and time-varying nature of the internal mechanisms of the reactor, it is inevitable that inaccurate model parameters will be used in the modeling process. These will lead to discrepancies between the modeled mechanisms and the actual reactor. In this paper, the difference is evaluated and shortened by means of neural network hybrid modeling. Based on the MATLAB/Simulink simulation platform, this paper firstly obtains the parameters that have the greatest influence on the linear model through sensitivity analysis and takes them as the object of neural network correction, then obtains the data required for offline training of neural network according to the mechanism model, linear model and the deviation of the two under different working condition levels, retains the neural network weights and thresholds obtained from the offline training, and finally utilizes the gradient descent algorithm to update the neural network weights and thresholds in real time in order to achieve the online calibration of the linear model. The final results show that the hybrid model can effectively reduce the steady-state deviation between the two models, which indicates that the hybrid modeling can effectively improve the accuracy of the established model and provide a solid foundation for the subsequent design of the control system based on the linear model.