An integrated learning risk prediction model for multisource data in nuclear power engineering

Abstract

For the modal matching problem of multi-source heterogeneous data from nuclear power plants, this paper proposes a risk warning system that integrates deep feature extraction and hybrid ensemble learning. A multi-modal deep autoencoder network is used to achieve high-dimensional feature alignment of cross-source data; a stacked architecture of GBDT and LSTM hybrid base learners is designed, and a dynamic weight adjustment strategy based on gradient optimization is introduced to enhance model synergy; an improved SMOTE sampling technique is adopted to alleviate the problem of sample imbalance. The main pump and pipeline system data of a certain nuclear power plant were measured, and the results showed that the model's F1-score was 12.4 percentage points higher than that of the traditional single model, the false alarm rate was lower than 1.8%, and the calculation delay met the requirements of real-time online monitoring.