Methodology Research On Condition Prognosis And Diagnosis Of Nuclear Power Plant Thermal-hydraulic System

The condition monitoring,prediction and fault diagnosis of nuclear power plant thermal-hydraulic systems can help operators to grasp the state of the plant and provide operational advice.These monitoring,prediction and fault diagnosis systems are of great significance to the nuclear power plant operational safety.The fast development of artificial intelligence algorithms based on neural networks in recent years provides many new approaches for condition prediction and fault diagnosis of nuclear power plant thermal-hydraulic systems.In this dissertation,a study was conducted on the condition prediction and fault diagnosis of nuclear power plant thermal-hydraulic systems based on neural network algorithms such as extreme learning machine(ELM),online sequential extreme learning machine(OS-ELM),radial basis function network(RBFN),recurrent neural network(RNN)and long short-term memory(LSTM)model.Natural circulation systems are widely applied in advanced nuclear power reactor design.However,complex and irregular flow instability phenomena are often observed in natural circulation systems.Therefore,the online monitoring of their operation condition becomes difficult.A short-term condition variable prediction model based on extreme learning machine neural network was proposed to solve this problem.The prediction model uses the condition data of thermal-hydraulic system as input and it is appropriate for the nuclear power plant thermal-hydraulic system with complex operation condition because of its extremely fast training speed.The input data structure of the prediction model was optimized using the genetic algorithm and the prediction accuracy was improved.A prediction model with online learning ability was also proposed based online sequential extreme learning machine(OS-ELM),which can further improve the prediction accuracy and reduce the overfitting brought by increasing hidden node number of extreme learning machine.The proposed condition variable short-term prediction model was also applied in condition variable soft sensing and sensor fault monitoring of thermal-hydraulic systems.The simulation experiments of the condition variable soft sensing model and the sensor fault monitoring model showed good results and they can be used as support for the online monitoring of nuclear power plants.A novel nuclear power plant thermal-hydraulic system fault condition diagnosis method was proposed based on improved radial basis function network.The proposed diagnosis method can identify the fault type included in the training set with high accuracy,and also has the ability of giving correct “don't know” diagnosis response to the fault transients not learned by the diagnosis model.The ability of giving “don't know” diagnosis response is not owned by other widely used neural network model.The proposed radial basis function network diagnosis model utilizes a self-organizing method based on K-means clustering to assign hidden basis function center,which reduces the number of hidden nodes.Furthermore,a new method of determining basis function variances for nuclear power plant fault diagnosis was proposed according to the characteristics of fault transient data,which can improve the diagnosis accuracy.The input data structure of the proposed diagnosis model was optimized using the genetic algorithm.In the optimization,a novel method of setting zoom factors for input data was proposed and the performance of the model was further improved.A prediction model of the remaining time to the reactor trip was developed based on deep learning algorithm RNN and LSTM,which were widely used in the field of artificial intelligence in recent years.The proposed prediction model can effectively learn the information from thermal-hydraulic system condition variable data sequence and shows better performance than the conventional feedforward neural network models using data from single time point as input.The LSTM node was used in the prediction model,so the vanishing gradient problem of simple RNN was solved and the accuracy of prediction was improved.The overfitting in the prediction model is avoided by adding a dropout layer to the prediction model of remaining time to the reactor trip based on LSTM recurrent neural network.A prediction model of important nuclear power plant thermal-hydraulic system condition variables during fault transients was developed based on LSTM recurrent neural network.The prediction model uses the condition data during the initial state of fault transients as input,and the model can provide long-term changing trend forecast for condition variables.A feature selection optimization method based on maximal information coefficient(MIC)was proposed for the prediction model.The feature selection using this method improved the accuracy of the LSTM recurrent neural network prediction model.