Data-driven fault diagnosis in nuclear power plant(NPP) utilizing CNN-based Bi-LSTM models

摘要: Accurate and long-term prediction of reactor parameters during severe accidents is critical for emergency response and accident management in Nuclear Power Plants (NPPs). However, existing data-driven methods are often hindered by error accumulation in long-sequence forecasting and fail to account for dynamic external interventions (e.g., operator actions). To address these challenges, this study proposes an integrated risk prediction framework. The framework first employs a dual-stage deep network for early fault classification. Subsequently, a prognostic model utilizing a CNN-BiLSTM-Attention architecture is introduced. A novel autoregressive injection strategy is developed to dynamically integrate secondary control signals—such as valve actions and safety injection status—into the iterative prediction loop, thereby stabilizing long-term inference and ensuring consistency with system operations. The model was trained and validated on a dataset generated by RELAP5 simulations of an M310 reactor under various accident scenarios (e.g., LOCA). Experimental results demonstrate that the proposed framework significantly outperforms standard CNN and LSTM baselines. It achieves high-fidelity predictions for key parameters (e.g., core outlet temperature, system pressure) over an extended 14,400 s horizon, while maintaining robust early fault classification capabilities. This approach provides a reliable, physics-aware decision-support tool for operator intervention during accident progression.