SELECTING FEATURES FOR INTEGRATED PRESSURIZED WATER REACTOR’S BEHAVIOR PREDICTION USING MACHINE LEARNING

抄録

The integrated pressurized water reactor (IPWR) is a nonlinear, large-lag, and strongly coupled system with hundreds of measured parameters. The correlation between various parameters can be divided into strong correlation, weak correlation and irrelevance. Selecting the suitable number and type of strong correlation parameters to represent the problem is the main challenge to study system behavior prediction. Two different schemes for simultaneously performing the selecting feature and the training of predictive model of system behavior based on back propagation neural network (BPNN) are presented. The first scheme uses the covariance method (CM) to select parameters with large correlation coefficient as features. The second uses principal component analysis (PCA) method to reduce dimensions of raw data, the data with lower dimensions incorporates information of all measurement data. PCA has high selecting feature performances with low numbers of parameters. The measured data was obtained from the integrated pressurized water reactor RELAP5 program, then the data was standardized. After selecting features of two schemes respectively, the features are trained by the same prediction model. The effectiveness of the proposed selecting feature methods is tested on the integrated pressurized water reactor RELAP5 program. The Minimum Departure from Nucleate Boiling Ratio (MDNBR) of reactor core is predicted under load tracking conditions, the results show that as input data of the predictive model of system behavior, the data generated by the PCA method has lower dimensions, but has the same superior prediction performance.