Pressurizer Water Level Reconstruction for Nuclear Power Plant Based on GRU

Abstract

Pressurizer water level is a vital parameter for marine pressurized water reactors (PWR). Operators use it as key evidence to comprehend reactors' operating conditions and identify their transients. Aiming at problems such as false water level and water level depression due to measurement faults, this paper applied deep learning techniques to the reconstruction on nuclear facilities' parameters. Considering the timing features of PWR operation, a reconstruction model for pressurizer water level were developed upon Gated Recurrent Unit(GRU) recurrent neural network theory. Six parameters highly coupled with pressurizer water level, including the average temperatures of reactor inlet and outlet, the pressure and temperature of pressurizer, the main coolant flow in primary circuit and reactor power, were selected as characteristic input parameters for the reconstruction model. By collecting the operating data of a marine PWR simulator whose reactor power transiently increased from 30% to 90%, the reconstructive pressurizer water levels were tested. The findings showed that the Mean Absolute Error(MAE) resulted from the GRU reconstruction model 0.003345. Compared with BP neural network models, the GRU model was not only able to more accurately approximate the true water level value, but also demonstrated better robustness, stability and convergence rate.