Multistep forecasting of state variables in nuclear power plants using deep learning

Abstract

In Nuclear Power Plants (NPPs), most monitoring and diagnostic systems operate based on the principle of Detection and Response (D&R), in which operator actions are triggered only after an anomaly is detected. While effective for real-time monitoring, this approach lacks predictive capability, which is critical for anticipating the evolution of accidents and enhancing operational safety. To address this limitation, this study investigates the use of Deep Learning models for multi-horizon forecasting the temporal behavior of key state variables during normal operation and postulated accident scenarios in nuclear reactors. Two datasets were employed: the LABIHS dataset, composed of simulated time series from a Pressurized Water Reactor (PWR) under a Loss-of-Coolant Accident (LOCA), and the SICA dataset, which contains real operational data from the Angra 1 nuclear power plant. The methodology included data preprocessing and data augmentation using instrumentation noise. Four deep learning architectures were evaluated: Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), Time-series Dense Encoder (TiDE), and Neural Hierarchical Interpolation for Time Series (N-HiTS). These models were trained using a sliding window approach and evaluated across multiple forecasting horizons. Comparative results showed that TCN outperformed LSTM among the classical models, while TiDE and N-HiTS achieved the best overall accuracy and stability across all forecasting horizons. With average MAE values of 1.01 ± 2.39 (LABIHS) and 1.45 ± 1.33 (SICA), these findings confirm the effectiveness of modern Deep Learning architectures for predictive monitoring in nuclear power plant operations.