Disruption prediction using a full convolutional neural network on EAST

In this study, a long short-term memory (LSTM) model is trained on a large disruption warning database to predict the disruption on EAST tokomak. To compare the performance of the proposed model with the previously reported full convolutional neural network (CNN) (Guo et al 2020 Plasma Phys. Control. Fusion 63 025008), the same data set and diagnostic signals are used. Based on the test set, the area under the receiver operating characteristic curve, i.e. the AUC value of the LSTM model is obtained as 0.87, and the true positive rate (TPR) is sim87.5%, while the false positive rate (FPR) is sim15.1%. Since the LSTM model is more sensitive to radiation fluctuations than CNN, the prediction performance of LSTM model is inferior to that of CNN model (for CNN, AUC sim 0.92, TPR sim 87.5%, FPR sim 6.1%). However, the advance warning time of LSTM model is 14 ms earlier than that of CNN. To reduce the FPR and improve the performance of the model, more fast bolometer channels are added as the input signals of the LSTM model, including the radiation from the upper and lower edges and the plasma core. Consequently, for the same test set, the AUC value increases to 0.89, and the FPR decreases to sim9.4%, but the TPR also decreases to sim83.9%. In addition, the sensitivity of the model to radiation fluctuations caused by impurity behavior decreases significantly, and the warning time becomes 8.7 ms earlier as compared to that of the original model. Overall, it is proved that deep learning algorithms exhibit immense application potential in the disruption prediction of long-pulse fusion devices.本研究基于大型等离子体破裂预警数据库训练长短期记忆网络（long short-term memory, LSTM），以预测EAST托卡马克（Tokamak）的等离子体破裂事件。为将所提模型的性能与已报道的全卷积神经网络（full convolutional neural network, CNN，Guo等2020 Plasma Phys. Control. Fusion 63 025008）进行对比，本研究采用了相同的数据集与诊断信号。基于测试集，长短期记忆网络模型的受试者工作特征曲线下面积（AUC）为0.87，真阳性率（TPR）约为87.5%，假阳性率（FPR）约为15.1%。由于长短期记忆网络模型对辐射波动的敏感性高于卷积神经网络，其预测性能逊于卷积神经网络模型（卷积神经网络的AUC约为0.92，TPR约为87.5%，FPR约为6.1%）。但长短期记忆网络模型的预警时间较卷积神经网络提前14 ms。为降低假阳性率并提升模型性能，本研究新增更多快速热辐射计（bolometer）通道作为长短期记忆网络模型的输入信号，涵盖等离子体上下边缘与芯部的辐射强度。据此，在同一测试集上，模型的AUC提升至0.89，FPR降至约9.4%，但TPR亦降至约83.9%。此外，模型对杂质行为引发的辐射波动的敏感性显著降低，预警时间较原始模型进一步提前8.7 ms。综上，本研究证实深度学习算法在长脉冲聚变装置的等离子体破裂预测领域具备巨大的应用潜力。