The Implementation And Verification Of An Intelligent Control Model For Steam Generator Level In Real-Time Simulation Scenario

As one of the major equipments in nuclear power plants,the steam generator（SG）plays an important role for the primary heat transfer loop.The control performance of steam generator level directly affects the safe operation of nuclear power plants.At present,PID controller is widely used for SG level control in pressurized water reactors.However,its fixed-parameter method lacks the features of optimization,self-adaptation,self-learning and other functions.As a nonlinearly variant system,the performance of steam generator level control system directly affects the power generation efficiency and even the safe operation of nuclear power plants.In order to better solve the problem of PID controller in level control,this thesis adopts the data-driven idea to optimize the steam generator level control to improve the control performance.The main work includes:1.The steam generator level control model in SIMULINK environment is constructed,in which the real-time data of different operating conditions are collected through the sim-ulation process.The collected data are used for the subsequent training and testing of time series prediction model based on deep learning methods.2.The zigmoid function is applied to LSTM to enhance the long-term ability of LSTM,generating a new type of recurrent neural network called zLSTM.The zLSTM model is combined with two-layer fully connected（FC）network to carry out feature extraction and dimension transformation for time series data.The test examples show that the mean square error of this model can reach the order of magnitude 10^-6in multivariate time series predic-tion tasks,which proves that zLSTM can not only respond to liquid level changes quickly but also output accurate prediction results faster.3.In this thesis,a TCP real-time communication system is designed and implemented,which synchronously transmits the level information from the real-time PID system to zL-STM model running in a deep learning framework for accurate predictions.The predicted level values are synchronously transmitted back to the control system,so that the level con-trol system can realize a forward-looking control scheme shorten the controller adjustment time to 2.435 seconds,and reduce the overshot by 29.8%.Compared with the traditional PID controller,its anti-interference performance is effectively improved on the premise of better control performance.The intelligent control model designed in this paper uses the real-time feedback data to generate a compensation input to the PID controller without modifying the control parame-ters.The optimization is realized by deep learning-based prediction from the zLSTM model that contains better the long-term memory.It provides a more suitable solution for multiple variate time series prediction in SG level control scenario and brings safer operation to the primary heat transfer loop of nuclear power plants.