Multi-step time series prediction for nuclear power plants based on variational mode decomposition and gated recurrent units

Abstract

Accurate prediction of reactor sensor data and fault diagnosis is critical for ensuring the safe and stable operation of nuclear reactors. It can effectively enhance reliability, reduce the risk of failure, and ensure the stable operation of nuclear power plants. However, the nuclear fission process and neutron flux vary over time, and physical quantities such as reactor temperature, pressure, and flow are influenced by factors like load, power regulation, and coolant flow, which result in nonlinear fluctuations. These fluctuations make sensor data complex and non-stationary, making it difficult for traditional methods to extract useful features effectively. To address this, we propose VMD-Bi-GRU, a novel hybrid model combining Variational Mode Decomposition (VMD) and Bayesian-optimized Bidirectional Gated Recurrent Units (Bi-GRU). VMD decomposes raw sensor signals into physically interpretable intrinsic mode functions (IMFs), effectively isolating noise and enhancing feature extraction. The Bi-GRU network then leverages its bidirectional temporal modeling capability for multi-step prediction. Crucially, Bayesian optimization automates hyperparameter tuning, maximizing model generalizability. Evaluated on CPR1000 simulator data, VMD-Bi-GRU achieved average RMSE reductions of 5.4 % and 55.2 % across 1–20 prediction time horizons for the first and second experimental datasets, respectively On the second experimental dataset, VMD-Bi-GRU reduces 20-step prediction MAE by 46 %. The reconstructed predictions are highly consistent with the original data ($R^2$ >0.7), enabling early anomaly detection and reflecting the reactor status with higher fidelity. This framework provides a reliable foundation for intelligent scheduling and predictive maintenance of nuclear power plants.