A hybrid BiLSTM-XGBoost model for short-term forecasting of key parameters in nuclear power systems

Abstract

Accurate prediction of operational state trajectories is the foundation of safe nuclear energy system operation, and precise transient parameter forecasting during Loss-of-Coolant Accidents (LOCA) is critical for early warning and risk mitigation. Conventional model-driven methods face a trade-off between computational cost and predictive fidelity, while single data-driven models are limited in modeling complex parameter couplings. This study proposes a hybrid BiLSTM-XGBoost prediction framework for short-term forecasting of key nuclear power system parameters. LOCA scenarios were simulated via the PCTRAN-PWR platform, with six core operational parameters sampled at 1-second intervals over 0–300 s to generate 1800 data points; the dataset was normalized and chronologically divided into training, validation and test subsets. After independent training of LSTM, CNN-LSTM, BiLSTM and XGBoost-BiLSTM models, their predictions were fused using an error-reciprocal weighting scheme. Experimental results show the hybrid model outperforms all standalone benchmark models significantly: the coefficient of determination (R²) for all six parameters exceeds 0.99 (nuclear power R² rises from 0.9460 of single BiLSTM to 0.9955, pressurizer pressure R² reaches 0.9994), with substantial reductions in MSE, RMSE and MAE. The model effectively captures both abrupt transient changes and gradual parameter trends in LOCA scenarios, providing a robust approach for nuclear power plant transient parameter prediction, critical technical support for nuclear power operational safety and intelligent supervisory control, and a generalizable solution for forecasting high-dimensional, time-dependent nonlinear parameters in other industrial systems.