Research On Algorithms For Transient Operation Analysis Of Nuclear Power Plant Based On Deep Learning

In the operation management of nuclear power plant,operators need to quickly and correctly judge the current operation state of nuclear reactor,and take a series of effective control measures in combination with operating procedures and their own experience.However,the experience of operators can not be used as a quantitative index,nor can they predict the impact of their operation of their operation behavior on the operation process of the nuclear power plant.Especially in the event of an accident,the possibility of further expansion of the accident caused by human error increases in a highly stressful environment for operators.Moreover,for beyond design basis or unknown accidents,the operating procedures have no expansion ability,and can not provide operators with fault type discrimination basis and accident decision-making scheme according to the current environment and nuclear power plant operation status.To solve the above problems,the algorithm of nuclear power plant transient operation state analysis based on deep learning is studied in this thesis.The argorithm has the functions of real-time calculation and prediction of transient operating parameters of nuclear power plant,identification of abnormal operating states and false alarm behaviors of nuclear power plant,and diagnosis of fault types.The algorithm can autonomously or assist operators to understand the operation status of nuclear reactor,and provide data support or judgment basis for operation control or accident disposal,and further improve the intelligent operation level of nuclear power plant control system.Firstly,the transient operating parameters prediction module can use this group of operating parameters to calculate or predict the change trend of one or several monitoring parameters at the next time.Then the abnormal operating status identification module can also use this group of operating parameters to identify whether the current nuclear power plant is in normal operation.If the transient operating parameters prediction module calculates that the current monitoring parameters are within the specified range,the abnormal operating status identification module considers that the current nuclear power plant is in normal operation.However,if the measuring instrument triggers an alarm,it indicates that the current instrument alarm signal is a false alarm.The above is the workflow of the false alarm identification module.When the abnormal operating status identification module considers that the current nuclear power plant is in abnormal operation,and the false alarm behavior identification module excludes the alarm signal as false alarm,it can be considered that an accident has occurred in the current nuclear power plant.Therefore,the fault diagnosis algorithm module can be used to judge the type of accident,so as to provide judgment basis for operators to formulate accident disposal plans.At the same time,in order to verify the influence of noise on machine learning algorithm,the consistency of the calculation results between transient operation data and simulation data part of this thesis proves that the difference between the two machine learning algorithm results is only related to the mean value of noise,when the machine learning algorithm is trained by using the simulation data and the transient operating data with noise,respectively.Therefore,it can be explained that when there is a lack of nuclear power plant transient operating data,the simulation data can be used to supplement the nuclear power plant transient operating data set without changing the final calculation results.For the transient operating parameters prediction module,the automated deep learning algorithm is composed of the core algorithm,quantum genetic algorithm and model agnostic meta learning algorithm.It can achieve high precision prediction of transient operating parameters in a short time,and the greatest advantage of the algorithm is that it can independently complete the design and training of the algorithm without the intervention of algorithm designers.The maximum error of transient operating parameters within 20s is less than 4%,and the prediction time is about 0.7 ms.At the same time,high precision prediction can be achieved within 5 s.The abnormal operating state identification module is composed of variational auto encoder and isolation forest algorithm.Its characteristic is that it can identify the abnormal operating state of nuclear power plant only by using the transient operating data during the normal operation of nuclear power plant,and the algorithm training does not need the accident condition data.Moreover,once an accident occurs,the abnormal state identification algorithm can determine the abnormal operating state of the nuclear power plant at the initial time of the accident.The false alarm behavior identification module is composed of three algorithms: BERT,sparse auto encoder,and isolation forest.The attention mechanism in the BERT algorithm can calculate the transient operating parameters in real time and determine other parameters that have the greatest impact on the calculation parameters.The sparse auto encoder and isolation forest algorithm can be used as redundant algorithms to improve the reliability of identifying the operating state of nuclear power plant.The false alarm identification algorithm can identify accidents with a correct rate of 100%,and identification time is about 0.13 s.The fault diagnosis algorithm module is composed of Ada Boost algorithm and XGBoost algorithm,which is characterized in that the two algorithms are the algorithm framework of integrated learning,and can replace the algorithm in the base learner to adapt to different fault diagnosis scenarios.Ada Boost is a binary classification algorithm,which is applicable to the fault diagnosis algorithm with a small amount of data,and the average recognition accuracy is 95%.The recognition accuracy of the accident 150 s after the accident can reach 100%.XGBoost is a multi-classification algorithm,which is suitable for algorithms with large amounts of data,and the average recognition accuracy is 73%.However,when the accident occurs 150 s later,the recognition accuracy can also reach 100%.Therefore,Ada Boost and XGBoost algorithms are reasonably selected according to the amount of data.To sum up,the nuclear power plant transient operating analysis algorithm based on deep learning can be used for nuclear power plant state identification and fault diagnosis,and can effectively enhance the safety,reliability and intelligence level of nuclear power plant operation.At the same time,the algorithm can also be used as an early verification scheme to provide design schemes and theoretical guidance for the design of other nuclear reactor intelligent control systems,and extand the application scope of the algorithm.