Disruption prediction on J-TEXT tokamak using ACO-BP-AdaBoost algorithm coupled with data augmentation

• Published in: The European physical journal. ST, Special topics Vol. 234; no. 13; pp. 3427 - 3439
• Main Authors: Lin, ZhiFang, Chen, TaiYuan, Yang, Yang, Yan, JuanJuan, Yan, Wei, Chen, ZhongYong, Zheng, Wei
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2025; Springer Nature B.V
• Subjects: Accuracy; Adaptive algorithms; Algorithms; Ant colony optimization; Atomic; Back propagation networks; Classical and Continuum Physics; Condensed Matter Physics; Data augmentation; Datasets; Deep learning; Disruption; Effectiveness; False alarms; Machine learning; Materials Science; Measurement Science and Instrumentation; Molecular; Neural networks; Optical and Plasma Physics; Physics; Physics and Astronomy; Plasma; Prediction models; Radiation; Random noise; Regular Article; Support vector machines; Tokamak devices
• ISSN: 1951-6355; 1951-6401
• DOI: 10.1140/epjs/s11734-025-01685-x

Accurate prediction of disruptions is essential for ensuring the safe operation of tokamaks. However, achieving high accuracy in data-driven disruption prediction models requires a substantial amount of experimental disruption data, which is not a feasible option for tokamaks, especially for future large devices. Data augmentation by adding Gaussian noise is an effective method to increase the dataset size, which can enhance the model’s robustness against various types of noise and improve the model’s generalization capabilities. Before data augmentation, the disruption prediction model of back propagation (BP) neural network is optimized by the ant colony optimization algorithm (ACO) and the adaptive boosting (AdaBoost) algorithm. The area under the receiver operating characteristic curve (AUC) achieves 0.9382 in this improved model. Based on the ACO-BP-AdaBoost model, different data augmentation strategies with various augmentation ratios are investigated. It is observed that data augmentation leads to an increment in AUC across all ratios. The best performance (AUC = 0.9677) of the model is obtained when the disruptive data are augmented fourfold and the non-disruptive data are doubled. Both the warning time and true positive rate are approaching the minimum requirements of ITER. Even when the training data size is decreased to 30%, the AUC of this model with data augmentation can be higher than 0.93, demonstrating the effectiveness of data augmentation in performance improving under relatively small target samples. However, with only 10% of the training data size, the performance of this improved model decreases significantly, which may be due to insufficient disruptive targets to learn features for prediction. Although the performance of this ACO-BP-AdaBoost model coupled with data augmentation does not completely satisfy ITER requirements, it provides a potential to expand the database by adding Gaussian noise, which may be helpful for disruption prediction.