A recurrent neural network for modeling natural circulation density wave instabilities

Advancements in machine learning and deep learning capabilities over several decades have resulted in models specialized for unique data structures and use cases. One such model, the recurrent neural network (RNN), is a deep learning architecture that incorporates temporal memory in its training. This is primarily performed utilizing long short-term memory (LSTM). Compared to dense neural networks (DNN), this feature gives the potential to model highly dynamic behaviors that would otherwise be limited using conventional machine learning models. LSTM is used here to predict two-phase flow instability, specifically density wave oscillation (DWO), which is characterized by significant oscillations in a boiling channel that can severely impact the safety of boiling water reactors (BWRs). Density waves produced by dynamic voiding in the channel dominate the system and maintain high amplitude flow oscillations which can cause cycles of dryout and rewet. In this paper, the RNN model is trained on experimental DWO data in order to test the abilities of the LSTM framework. Training data is obtained from tests using a single channel 10 × 10 BWR bundle geometry with electrically heated rods at 7–8 MPa. The model is tested with different architectures and utilizes different optimization methodologies to determine the impact on results. • Recurrent neural networks were trained using density wave instability data to determine predictive capabilities • LSTM-based RNN performance showed good results for modeling oscillation behavior • Parametric investigation provided insight on model optimization for this dataset