iLMformer: A data-driven hybrid architecture for multivariate time series parameter forecasting of nuclear engineering systems

Nuclear energy, which is considered a strategic low-carbon power source, has become increasingly important in the context of the current global effort to cut carbon emissions. However, nuclear power plants still face many challenges, such as complex dynamic characteristics and the fact that operators can only monitor a small number of changes in critical reactor operating parameters (including the reactor loop temperature, pressure, flow rate, etc.) at a time when an accident occurs. These issues significantly reduce nuclear system safety. This study presented the specific iLMformer hybrid algorithm architecture, which combined the iTransformer architecture with long short-term memory networks to overcome the shortcomings of current data-driven algorithm models, specifically their inadequate long-term predictive accuracy and poor noise resilience. The model's applicability was confirmed through the utilization of multidimensional time-series datasets obtained from actual operational data collected via a system-level thermal-hydraulic test bench. After that, the model's ability to handle noise and generalize was tested using perturbation data made by RELAP5, which included Gaussian white noise. Performance tests indicate that iLMformer works better than convolutional neural network and long short-term memory network baseline models and converges faster than standard Transformer models. Tests of the device's ability to handle noise showed that it is forceful. These results show that iLMformer's feature fusion and noise adaptation mechanisms work well, making it a very accurate and reliable way to predict operational parameters in smart nuclear power plants.