Demonstration of combined reduced order model and deep neural network for emulation of a time-dependent reactor transient

In this paper, a Reduced-Order Model (ROM) was constructed to approximate nuclear reactor neutronic transient behaviour. The ROM was developed by first using the Principal-Component Analysis (PCA) to reduce the output space of the response of interest, i.e, the time-dependent reactor power. A Deep-Neural Network (DNN) is then employed to build a surrogate model that maps the model inputs, i.e, the kinetic data, to the reduced-order output. While many previous applications of machine learning have focused on using algorithms to generate the response of singular Figures of Merit (such as peak clad or peak fuel temperatures over an entire transient) relatively few works have developed methods that can replicate the transient output of computer codes over the duration of the simulations. The difficulty in using machine learning to reconstruct the entire transient response is that, given the duration of the transient and the relatively fine time steps used, the output space can be very large thus making standard machine learning methods difficult to apply. In this work, PCA is first used to identify the key components of the output responses such that DNN can then be used to map the input parameters to the Principal component time-dependent behaviour. Then the neural network can reconstruct these components and emulate the output transient response based on the input parameters selected. Thus the methodology can provide time-dependent output results rather than single figures of merit. In this work, the LRA benchmark was used as a demonstration problem. The Detran time-dependent diffusion solver was used to generate the training and test data for a power control-rod ejection scenario. The results show that the combined ROM-DNN methodology is able to reconstruct the output power trajectory with high accuracy and with significantly reduced computational overhead.