Work

Time-series prediction of electromotive force for dissolved oxygen in lead-bismuth eutectic based on hybrid neural network

Year: 2025

Type: article

Abstract: Dissolved oxygen concentration is critical for the resistance of structural material corrosion in lead-cooled fast reactor (LFR). To improve the prediction accuracy of dissolved oxygen change trend, a novel forecasting approach, named Feature-Enhanced Variant Attention Network (FEVA-Net), based on hybrid neural network, is proposed, combined with temperature as the core feature related variable. In this paper, based on the experimental data acquired in the liquid lead-bismuth eutectic (LBE) loop facility, KYLIN-II-M, a prediction model for electromotive force of CNN-BiLSTM integrated with a variant attention mechanism is constructed. Firstly, the experimental data is processed by Wavelet Adaptive Soft Threshold Denoising (WASTD). Subsequently, the temperature sequence is decomposed by Empirical Mode Decomposition with Pearson Correlation Coefficients (EMD-PCCs), then the relevant modal components are obtained and input into CNN-VA for feature extraction. Meanwhile, the historical data of temperature and electromotive force representing oxygen concentration are input into BiLSTM-VA for feature extraction. Finally, the prediction output is obtained through the fusion of the linear layer. Experimental results demonstrate the advantages of the FEVA-Net compared with EMD-CNNBiLSTM model in terms of the RMSE, MAE, MAPE and R2, which are improved by an average of 31.40%, 43.58%, 2.05% and 4.17% respectively. This study illustrates the promising accuracy of the proposed model, especially for dissolved oxygen process in practice.

Source: Nuclear Engineering and Technology

DOI: 10.1016/j.net.2025.103967