A control-oriented hybrid modeling method for small pressurized water reactors based on linear models and neural networks

Abstract

Linear models such as state-space and transfer function models are widely used in the control design of nuclear reactors. However, due to the simplifications in model linearization, parameter and structural errors inevitably exist in linear reactor models. This paper proposes a neural network-based hybrid modeling method for small pressurized water reactors (SPWRs) to calibrate their inaccurate linear models. First, key parameters that have significant impacts on the SPWR linear models are obtained through sensitivity analysis. Then, the gradient descent algorithm is applied for the neural networks training with the loss functions constructed with the two most important control objectives of the SPWR. Finally, the neural networks are used to calibrate the selected key parameters in the SPWR linear models to construct accurate control-oriented hybrid models. Simulation results of the SPWR show that the neural network-based hybrid modeling method can significantly decrease the transient and steady-state errors of the linear models.