The Study Of Plasma Disruption Prediction Based On Deep Learning On EAST Tokamak

The magnetically confined tokamak is one of the most promising devices in the research of controlled fusion.Presently,there are many problems in the tokamak to overcome,plasma disruption is one of them.Plasma disruption is caused by the rapid development of magnetic fluid dynamics(MHD)instability events to uncontrollable,resulting in the decrease of plasma confinement performance,rapid loss of energy storage,and rapid extinction of plasma current.When the plasma is disrupted,a large amount of energy(thermal energy and magnetic energy)is deposited to the first wall material or other components of the device in the form of thermal flux,halo current and runaway electrons in a very short time,which can cause damage to the device for a long time.With the increase of the size of fusion device and the improvement of operation parameters,such as ITER and CFETR,plasma disruption will be more harmful to the device.Therefore,early warning of disruption and effective disruption avoidance or mitigation are very important for the safety of the device.Since the complexity,nonlinearity and rapacity of plasma disruption development,the understanding of the physical mechanism of plasma disruption is not very clear,and it is difficult to predict plasma disruption based on the model of plasma disruption mechanism.For the past 20 or 30 years,plasma disruption warning has been a technical problem to be solved by various magnetically confined tokamak devices.With the development of big data and the improvement of computer computing ability,datadriven plasma disruption early-warning research gradually shows its advantages,which can achieve high accuracy and have a long time of early warning.Based on the EAST superconducting tokamak,after understanding the characteristics of plasma disruption in detail,this paper classifies and counts the disruption types.On this basis,convolutional neural network and long short-term memory network are used to train and learn disruptive data and non-disruptive data,and two sets of deep learning-based dsiruption warning models are built successfully Finally,after comparing the two algorithms,a hybrid deep neural network algorithm is constructed by combining the advantages of the two algorithms.Before the occurrence of plasma disruption,there will be changes in certain physical parameters.These precursor parameters can be used as input parameters for disruption warning.Firstly,the disruption discharges in the last six years is analyzed and counted in this paper.We found that in the EAST Tokamak experiment,more than 50%of the disruptions came from impurity sputtering.The impurities mainly came from the light impurities in the first wall,the heavy impurities in the divertor components and the metal impurities at the wave antenna port.Secondly,the disruption types are vertical displacement instability and density limit disruption.After determining the type of disruption on EAST,the physical parameters,such as magnetic signal and radiation signal,which are highly correlated with the development trend of disruption are selected,and the disruption warning database is built on EAST.Based on the disruption warning database,researches on the disruption warning models of two deep learning algorithms are carried out.The disruption warning model based on full convolutional neural network is used to train and learn the diagnosis signal data.The results show that the area under receiver operating characteristic curve(AUC)of the disruption warning model based on the full convolutional neural network is 0.92,which can accurately warn 87.5%of the disruption discharges,6.1%of the safety discharges error warning,and the average warning time is 46 ms.On the basis of the same data set and diagnostic signal,the long short-term memory network is used for training and learning.In the same offline test set,the AUC value of the disruption warning model based on the long short-term memory network is 0.87,the accurate warning rate of disruption discharges is 87.5%,the misjudgment rate of safety discharges is as high as 15.1%,and the average warning time is 60 ms.From the analysis of the results,it is found that the convolutional neural network has a high tolerance to the instantaneous abnormal changes of the input signal,and has obvious advantages in the extraction of disruption features,while the long short-term memory network model has obvious advantages in the early warning time.In this paper,the hybrid neural network model framework is constructed by combining the advantages of the two algorithms and the disruption warning database is expanded.Based on MHD disturbance array signals and radiation array signals,the two-dimensional convolution layer is used to extract the array information to construct one-dimensional effective features,and the remaining diagnostic signals are extracted through parallel one-dimensional convolution layer to extract their one-dimensional effective features.All the extracted effective features are fused,and the long short-term memory network algorithm is finally used for training and learning.The test results show that the AUC value of the hybrid neural network disruption warning model reaches 0.95,which can accurately warn 90.9%of disruption discharges,and only 2.4%of safe discharges are false alarms.