A preliminary study on activated corrosion product source term prediction in pressurized water reactor using recurrent neural network

Abstract

The activated corrosion products are major source of the collective effective dose in the maintenance and repair work, and various mechanistic models have been developed to evaluate its formation, transportation, and deposition process. Although many mechanistic models have been developed for activated corrosion product source term prediction, there is a drawback that the mechanistic model cannot fully utilize the historical data. In this study, an activated corrosion product source term prediction model is developed based on RNN, including classic RNN, LSTM, and NARX. Two different prediction algorithms are proposed and the performance of algorithms with different network structure is evaluated for one time-step and multiple time-steps. According to the results, RNN is a promising method for activated corrosion product source term prediction.