A team of scientists from the Institute of Nuclear Energy Safety Technology (INEST), the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, has developed a novel artificial intelligence framework that could improve the way we forecast the internal conditions of nuclear power plants.

The results were recently published in the journal Energy.

In nuclear power systems, small warning signs of equipment failure—like subtle temperature or pressure fluctuations—often appear long before major problems occur. Detecting these early signals is crucial for avoiding accidents. However, making reliable long-term predictions about key plant conditions has been a major challenge due to the complex physics involved, including strong nonlinear behavior and the interplay of multiple physical processes.

In this study, the team introduced a new three-stage AI framework- AFDI-Net, short for Adaptive Feature Decomposition and Interaction Network. The method breaks down the prediction process into three parts: decomposition, reconstruction, and forecasting.

In the first step, AFDI-Net uses a smart decomposition technique to untangle complex time-series data into clearer patterns based on trends and cycles. This helps make sense of the raw operational data. Then, in the second stage, the system reconstructs these patterns while learning both short-term details and long-range relationships within the data. Finally, a straightforward forecasting layer turns these learned patterns into predictions of future system behavior.

Beyond this new algorithm, the team also proposed a new way to evaluate how well these prediction models perform—not just by measuring accuracy, but by assessing how useful the predictions are for practical downstream tasks, such as plant safety assessments or equipment maintenance planning.

According to the team, AFDI-Net consistently outperformed traditional prediction methods in both short- and long-term forecasting tasks, showing particular strength in capturing the intricate dynamics of nuclear power systems.

Article: https://doi.org/10.1016/j.energy.2025.135784

Framework and algorithm structure of the AFDI-Net in nuclear power systems.