A Novel Unsupervised Gated Recurrent Neural Network for Plasma Disruption Prediction in Aditya Tokamak Using Dynamic Threshold-Based Temporal Differentiation

Abstract

Plasma disruption prediction is essential for sustaining stable nuclear fusion reactions. Existing data-driven approaches face limitations due to their dependence on labeled datasets, which are often difficult to curate in dynamic plasma environments. Also, these models typically rely on setting a fixed threshold—a manually defined cutoff point to detect fluctuations in plasma current that may indicate an impending disruption. This threshold is manually defined and remains constant, which can make it ineffective under evolving plasma conditions, where the nature of fluctuations may change over time. To address the limitations, this study proposes an unsupervised Gated Recurrent Neural Network model with a Dynamic Threshold-based Temporal Differentiation Algorithm (GRNN-DTTD) to predict disruptions. This threshold is formed by continuously analyzing temporal variations in plasma current fluctuations, allowing it to adjust based on evolving signal patterns. This adaptive mechanism enables the GRNN-DTTD to detect abnormal trends associated with impending disruptions without the need for pre-labeled training data. By learning directly from variations in the input signals over time, the model operates in an unsupervised manner, which identifies disruptive patterns and issues early warnings. Experimental evaluation was conducted on Aditya dataset (133 training shots, 91 unseen testing shots) which demonstrates the model’s effectiveness by achieving 98.9% prediction accuracy with warning times of 12–30 ms prior to disruption events. The results show that the proposed framework avoids manual threshold setting, eliminates dependency on labeled data, and improves adaptability to changing plasma conditions.