Machine learning-based water-level forecasting for marine steam generators under rolling conditions

In marine nuclear power plants, steam generator (SG) water level control is crucial for reactor safety, but it is challenged by two key factors: the transient inverse water level response to inventory changes under low-power conditions, and the superimposed influence of wind/wave-induced rolling motions. These factors induce nonlinear free surface fluctuations, rendering real-time measurement signals unreliable. While computational fluid dynamics (CFD) provides high-fidelity predictions, its high computational cost precludes real-time application. Conversely, traditional PID controllers struggle to handle these time-varying dynamics. This study employs Long Short-Term Memory (LSTM) and Backpropagation (BP) neural networks trained on CFD-simulated data to predict SG water levels under rolling conditions. The models leverage machine learning to capture sequential dependencies, mitigate measurement noise, and enable efficient real-time predictions. Validation results show that prediction errors are controlled within ±1.5 %, and the coefficient of determination (R²) exceeds 0.95. This confirms the superiority of neural networks in predicting SG water levels in complex marine environments.