Diagnosis And Prediction For Loss Of Coolant Accidents In Nuclear Power Plants Using Deep Learning Methods

Safety System is one of the crucial components of nuclear power plants.Malfunction of Safety System,will cause inestimable loss and irreversible damage.The current control methods are mainly based on the“Detection-Response”principle.Safety systems only act when the abnormal conditions have taken place,leading to a delayed response to conditions that may cause severe damages.Moreover,the traditional statistical methods cannot effectively solve the issue of non-linear process’s prediction in abnormal conditions.Through exploring and analyzing the massive historical data,the deep learning method emerging in recent years can grasp the operating rules of various operation conditions,predict the trend of real-time process and upgrade the safety control of nuclear power from“post-occurrence response”to“early diagnosis and intervention”.This method will greatly enhance the safety margin of nuclear power plants,and thoroughly work out the situation of“abnormal condition inevitably happening”.Therefore,this thesis analyzes the early diagnosis and prediction of LOCA in the reactor.Based on the machine learning and technology of deep learning,this thesis also focuses on the diagnosis of accident caused by coolant loss and the trend prediction after its occurrence.The research of this thesis is divided into four parts:（1）The fault mechanism of LOCA is analyzed.This thesis demonstrates the status of traditional research method.Through analyzing the disadvantages of traditional method of LOCA’s diagnosis,the thesis also puts forward a new diagnostic method of Att-Conv LSTM.The accuracy of diagnosis for five LOCA break seizes reaches 96%,which proves the validity of the proposed method.（2）The importance of trend prediction after the occurrence of LOCA is also analyzed.Aiming at the shortcomings of the existing LSTM-based prediction model,an improved CNN+LSTM+Res Net prediction model is proposed.For the predicted value of 60%of the 0.2cm² break size,the loss value is as low as the 1.936×10^-3,proving the effectiveness of the proposed method.（3）Aiming at the situation that the original simulation data set has few samples and is unlabeled,a script based on Python language is constructed to expand and label the data set.The break-size identification efficiency is effectively improved,so is the prediction accuracy.（4）Aiming at the situation that tremendous expert experience is needed for fault diagnosis,this thesis proposed to combine the diagnosis model of LOCA with the trend prediction model into a new model for the purpose of“diagnosis+prediction”.