Multi-step prediction of main pump leakage in nuclear power plants with an additive model

With the global demand of clean and low-carbon emission energy sources, safety in the nuclear power industry has gained widespread attention. The accurate and timely long-term fault prognostics of critical equipment in nuclear power plants (NPPs) helps to effectively schedule maintenance, thus reducing operation and maintenance costs while ensuring safety. Unpredictable future load and operations make long-term prediction of time series data in NPPs quite challenging. In this study, we proposed a framework for multi-step prediction based on an empirical mode decomposition (EMD) algorithm, which decomposed complex time series data into several components with simple trends. The useful signal and noise in the components were then divided by a relative ratio method. Candidate prediction models from the model library were trained and tuned on the signal-dominated components. The appropriate model was selected for prediction according to the characteristics of each component signal, and the prediction results were summed and smoothed with an additive model. The proposed method, noted as the denoised EMD method using an additive method (EMD-AM), was applied to the leakage prediction at shaft seals in the reactor coolant pump (RCP) of five different NPPs. Root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute scaled error (MASE) were selected to determine the prediction accuracy and uncertainty. Compared with EMD-AM and EMD method using long short-term memory (EMD-LSTM), the experimental results show that the denoised EMD-AM has the overall advantage of simplicity and accuracy.