Enhancing Nuclear Power Plant Operational Forecasting With Transformer Neural Networks: a Time-Series Data Approach

Abstract The forecasting of critical operational parameters within nuclear power plant is paramount for ensuring their operational safety and efficiency. In the realm of predictive analytics, the Transformer neural network architecture has distinguished itself by delivering superior performance across a diverse array of fields. Despite its proven efficacy, the exploration of Transformer models within the nuclear energy sector is notably sparse. This investigation seeks to bridge this gap by deploying Transformer-based predictive methodologies on a dataset characterizing the loss of offsite power scenario at a nuclear power plant. Through the application of the Tree-structured Parzen Estimator optimization algorithm, the study meticulously optimizes the neural network architecture to refine its predictive precision. The research effectively forecasts critical physical metrics such as core pressure vessel water level, core outlet temperature (average), hotleg temperature, coldleg temperature, steam line pressure, steam generator secondary feedwater flow rate. The empirical results of this study are promising, demonstrating the Transformer model’s good capability in predictions, with the normalized Root Mean Square Error for all evaluated physical parameters on the test dataset consistently remaining below 0.056. This outcome underscores the Transformer’s potential as a robust tool for predictive analytics in nuclear power plant operations.