Abstract
Steam generator water level is a complex thermal hydraulic system, which in-ludes nonlinearity, time-variant characteristic, non-minimun phase, mechanical limit and mesurement error. Particularly, “false water level” phenomenon caused by nonminimun phase greatly affects the control of steam generator water level. Thus, in order to greatly help PWRs operators control the water level, enhance the safety of PWRs operation and improve the economic of PWRs, it is of importance to study the technology on prediction of steam generator water level variation trend under normal operation condition of PWRs. A predictive model for U-tube steam generator (UTSG) is in need as an assistance for operators to better control the UTSG water level. The training data, i.e., the variation of UTSG water level in uenced by the sudden change of steam ow rate and feedwater rate, were obtained from a differential equation model, i.e., Ivring’s model. Long short-term memory (LSTM) and gated recurrent unit (GRU), as powerful time series prediction algorithms in deep learning eld, were used to train the UTSG prediction model. The result shows that the trained LSTM and GRU based predictive model have good performance on testing data set with both the one-step prediction and multi-step recurrent prediction. It can be used as a prototype of steam generator level prediction without the intervention of automatic control system.
