Development of Learning-based Model Predictive Control Framework for SMRs

Abstract

In this report, we document the development of a reactor dynamics and Learning-Based Model Predictive Control (LBMPC) algorithm for the autonomous reactivity control of a Small Modular Reactor (SMR). The reactor dynamics model includes the Point Kinetics Equations (PKE), Thermalhydraulics (TH) models, and Xenon dynamics. The position-dependent control rod worth is used to demonstrate a realistic situation. The nonlinearity of the reactor dynamics models causes a model mismatch with the linear state-space model used in the MPC controller, degrading the accuracy of the controller. The LBMPC controller is developed to minimize the error caused by the model mismatch. The Gaussian Process Regression (GPR) algorithm is used to train a way to update the state-space model as reactor condition evolves. In the training, the nonlinear model is successively linearized and the piecewise state-space model information is provided to the GPR. The trained GPR model provides improved state-space models to the MPC controller every time step resulting in better accuracy for reference power tracking