Prediction basis visualization method of plasma rupture predictor

Abstract

The invention belongs to the plasma control technology, and particularly relates to a prediction basis visualization method of a plasma rupture predictor. The method comprises the following steps: determining a neural network model for Tokamak plasma large rupture prediction, calculating a neural network model output result without disturbance, calculating a model output result under the disturbance condition, and after determining the sensitivity distribution value of the neural network input signal, carrying out normalization processing. The method is used in a fracture prediction task, the situation that algorithm prediction accuracy and interpretability need to be balanced does not exist, feedback can be provided for the Tokamak discharge experiment in real time, and the method can be used for Tokamak experiment correlation analysis to obtain the correlation between each physical parameter and fracture.

Description

technical field [0001] The invention belongs to plasma control technology, in particular to a method for visualizing the prediction basis of a plasma rupture predictor. Background technique [0002] Existing tokamak plasma large rupture prediction algorithms are usually based on traditional machine learning or neural network schemes. These algorithms all face the problem of high algorithm opacity and poor interpretability of the output results. At present, there are some methods for visual analysis of the results of these rupture prediction algorithms, which can be roughly divided into two categories in principle: (1) using some traditional machine learning models that are naturally interpretable; A certain input channel and do a comparative experiment to judge the importance of this input channel [0003] Naturally interpretable traditional machine learning models mainly include various integration methods of decision tree models, including but not limited to methods such ...

AI Technical Summary

Problems solved by technology

However, this method strictly stipulates the machine learning model used by the rupture prediction algorithm, and the tree model is not the best solution for pattern recognition at present, so this method is not adopted by most researchers.

[0004] The method of removing a certain input channel in the data set and doing a comparative experiment can only conduct a unified analysis on the importance of each input channel on the entire data set, but it is difficult to be specific to a certain discharge, so the visualization results of this type of method are usually only It can be used as an optimization analysis method in the algorithm development process, but cannot be the basis for actively avoiding major rupture events

Images