Abstract: In order to explore the establishment of the HL-2A/3 experimental drift wave mode database and take it as the sample data, through the machine learning method, the artificial neural network is used to predict the occurrence and intensity of drift wave mode instability on tokamak discharge and then offer the supporting conditions to realize the real-time control of the HL-2A/3 plasma. A basic database of low frequency drift wave modes is constructed based on the calculation results for the case of four primary parameters (h, e, q, s). Then, based on the back propagation neural network (BP network) and the support vector machine (SVM) models, the neural network modeling and programming experiments were carried out, respectively, which verified the feasibility of intelligent prediction of ion temperature gradient (ITG)\trapped electron modes (TEM) instability of HL-2A tokamak. The preliminary results show that by further expanding the database by including all the sensitive parameters, accelerating the training speed of BP network, or adopting more complex models such as deep learning, the ultimate target of drift wave mode prediction can be finally achieved.

Key words: HL-2A tokamak, Drift wave mode, Database, Machine learning, Artificial neural network, Feasibility study

摘要： 本文探索建立了HL-2A/3装置实验漂移波模数据库，并以此作为样本数据库，通过机器学习方法，利用人工神经网络预测托卡马克放电中漂移波模不稳定性的发生及其强度，为实现HL-2A/3等离子体实时参数控制提供参考。首先基于电子/离子温度梯度(h)、俘获电子份额(e)、局域安全因子q和磁剪切s等4个基本参数构成的参数数据组(h, e, q, s)作为变量，其他参数取有效的常数值，利用HD7代码计算相应模特征值数据，构建了一个低频漂移波模基本数据库。然后，基于BP神经网络与支持向量机(SVM)模型，分别进行了机器学习建模与编程实验，验证了对HL-2A装置离子温度梯度(ITG)\俘获电子模(TEM)不稳定性进行智能预测的可行性。研究结果表明，通过将参数集与数据集进一步扩充成完备数据库、并加快BP神经网络训练速度、或采用深度学习等更复杂模型，可以最终实现前述漂移波模预测目标。

关键词: HL-2A托卡马克, 漂移波模, 数据库, 机器学习, 人工神经网络, 可行性研究