Learning and ensemble based MPC with differential dynamic programming for nuclear power autonomous control

Abstract

Reactor autonomous control is one of the key research fields of advanced reactor R&D. Although some basic achievements have been made, it is still an unfinished goal to realize the intelligent autonomous control of the reactor. In this paper, some learning-based and ensemble-based predictive models and some model predictive control strategies combined with ensemble Kalman filter (EnKF) are proposed to realize a feasible path to the reactor autonomous control. For a typical point reactor and the operation target, reactor state space predictions are completed by mechanism-based models and data-driven learning-based models respectively. Mechanism models based on parameter perturbation and the neural network models that have higher accuracy are selected as predictors of model predictive control (MPC). The MPC results show that the model with less precision (e.g. support vector machine and random forest regression) has large oscillation results, while the predictive control based on neural network and mechanism model can effectively achieve the control objectives. To improve the accuracy of MPC, ensemble learning methodologies with a second learner based on linear regression, stacking different predictive models into one model are proposed. The accuracy of MPC with the ensemble model can exceed or be close to the optimal individual results. EnKF is proposed to estimate core state and covariance, which would be helpful to restrain the influence of observation uncertainties and predictive model error. In addition, some limitation of the proposed methodologies is discussed and needed further study.