A preliminary study of digital twin for nuclear reactor dynamics: a synergy of machine learning and model predictive control

The operational stability of small modular reactor(SMR) in unmanned environments is a critical factor restricting the application of nuclear power. In these conditions, reactor faces the issue of sparse communication data, making it difficult to accurately grasp reactor's operational state. Digital twin for reactor is an effective solution to this problem. However, the complexity of reactor prevents straightforward synchronization of digital model's state with sensor measurements from physical system. Therefore, this paper proposes a method to achieve digital twin of reactor, comprising a data-driven state synchronization method and an online calibration method based on model predictive control (MPC). This approach combines hybrid dilated convolution(HDC) and long shortterm memory (LSTM) to enable initial state synchronization. Online calibration utilizes point-kinetics model and core thermal-hydraulic model as predictive models.Controllable characteristic calibration parameters are selected as control variables, and a set of control variables that minimize model calculation bias is determined using the particle swarm optimization algorithm(PSO). These control variables are used to predict the transient changes in reactor. Simulation results indicate that the maximum mean absolute error (MAE) of state synchronization remains below 1.5. The error from online calibration also meets engineering requirements, and a lyapunov stability analysis further verifies the system's stability.Thus, the method has potential application value in other industrial processes with complex dynamic behaviors and can be extended to systems with similar characteristics.