Abstract

During abnormal conditions of nuclear power plants that do not lead to automatic scram immediately, operators need to monitor the transient condition of nuclear power plant in order to evaluate the development rate of the abnormal conditions. Conventional condition monitoring methods cannot provide operators with an alert or forecast for automatic scram. In this work, a deep-learning model is presented for predicting the automatic scram assuming no operator intervention. The presented prediction model is based on Long Short-Term Memory (LSTM) model, which is a special type of gated Recurrent Neural Network (RNN) designed to adaptive control the memory length of the learned features and showing good performance in Sequence to Sequence (Seq2Seq) problems. The predicting model is trained by transient data of abnormal conditions: the input of model is the nuclear power plant monitoring data; and the output is the remaining time from current time to reactor trip. The predicted remaining time to reactor trip decreases with the development of abnormal condition, thus the output of prediction model will generate a countdown to reactor trip. Prediction experiments were conducted on simulated abnormal condition data generated by a full-scale simulator for a 300MW pressurized water reactor (PWR). The data with artificial noise were also used to test the performance of the presented model. The experiment results shows that the presented prediction model has better prediction performance and is more robust to noisy input than models based on one-hidden-layer feedforward neural networks or simple RNNs (e.g., Elman network). The presented intelligent prediction model can provide nuclear power plant operators with a countdown alert about the upcoming reactor scram, as such the operators can prepare or take countermeasures for it.

This content is only available via PDF.
You do not currently have access to this content.