Temperature forecast for spent nuclear fuels (SNFs) in dry storage using time series decomposition and machine learning

Abstract

The spent nuclear fuel (SNF) is a high-level radioactive nuclear waste disposed from the nuclear power plants. Currently, SNF is stored in dry cask storage systems (DCSSs) for extended interim dry storage. The thermal profile of DCSSs during their dry storage has been identified as the top research priority. This study developed a temperature forecast framework for SNF in dry cask to predict their temperature variation over time. A framework was developed with six steps: data collection, data cleaning, sensor aggregation, time series decomposition, model training and evaluation, and long-term forecast with spatial temperature variability. The proposed framework was validated using real thermal data collected from M5 and Zirlo cladding materials. Research results found that the time series decomposition plays a key role by extracting the trend component from the original time series. Machine learning (ML) models trained on decomposed data achieve smaller errors compared to models trained on the original data. Among the four models evaluated, autoregressive integrated moving average (SARIMA) outperforms other models and has the best robustness across different training-testing split windows. Long-term temperature forecast demonstrates a gradual cooling trend and reduced temperature variability across fuel locations to the year of 2050. Conclusions and methodologies from this study are promising to support the decision-making for extended dry storage for high-level radioactive wastes.