Accurately predicting key operating parameters of the nuclear power plant (NPP) is crucial for the safety and efficiency, but the complexity and variability of reactor systems create high-dimensional, strongly correlated data that traditional small-scale models often fail to capture.

Recently, researchers from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences developed NPP-GPT, an artificial intelligence framework that uses pre-trained large language models to accurately forecast NPP operating parameters over the long term. The study was published in Applied Energy.

The NPP-GPT framework first aligns numerical time-series data with the representation space of a pre-trained language model through input embedding reconstruction and self-supervised learning with random masking. Then, domain knowledge is incorporated via LoRA (parameter-efficient fine-tuning) by adjusting the Q/V projections in GPT-2' s self-attention modules.

NPP-GPT adopts a cross-modal transfer learning strategy which preserves the general capabilities of the pre-trained model while enhancing forecasting performance and maintaining training and deployment efficiency.

Tests showed that NPP-GPT performed exceptionally well across six typical operating-condition datasets. In multivariate, multi-step forecasting tasks, it outperformed several mainstream time-series methods, maintaining high accuracy even as the forecasting horizon extended. Evaluations under cross-condition transfer, noise interference, and missing-data scenarios demonstrated its robustness and generalization.

This framework provides more reliable predictive information for online safety monitoring and operational decision support in NPPs, and opens up new possibilities for applying large language models in the nuclear energy field.