DEVELOPMENT OF ESTIMATION METHOD USING MACHINE LEARNING TO EVALUATE BREAK DIAMETER IN PWR LOSS-OF-COOLANT ACCIDENT

抄録

This paper describes development of a method based on machine learning for break diameter estimation in PWR loss-ofcoolant accidents (LOCAs). This estimation uses the data from a safety parameter display system (SPDS) that are transferred from power stations to the Japanese nuclear regulatory body. Two features extracted from the reactor coolant pressure p were used for learning: the minimum pressure decreasing rate dPmin before p becomes less than 4 MPa after the occurrence of the LOCA, and the time τ_dPmin between the point when p equals the saturated vapor pressure and the point when dPmin was obtained. The programming language MATLAB was used to extract these features from the SPDS data and to learn them using a support vector machine (SVM) with the 2nd order polynomial kernel function. There were several factors that cause deviations in the feature extraction results such as break position in the reactor coolant piping and the one-minute sampling timing of the SPDS. After data were simulated by the severe accident code MAAP4 and one-minute sampled, the input break diameter D was learned by the SVM with these features extracted from the simulated data. The learning result D* showed a relative error of plus or minus 13% and standard deviation of 5.6% under the deviation by break position and sampling error. It was confirmed that these learning results were valid even if the time interval from the reactor shutdown to the occurrence of the LOCA was changed from 0 to 30 or 60 minutes. It was judged that the learning results can be put to practical use.