Time series classification algorithms for confinement regime identification in C-Mod

This work addresses the problem of confinement regime identification as a multivariate time series classification task. We present results demonstrating that using feature extraction techniques significantly improves the performance of time-independent classification models [1], achieving accuracy comparable to more complex deep learning approaches [2,3]. For example, adding lag and shifted features to time independent models (random forest) improves the scores (Cohen’s Kappa) from 0.76 to 0.87 for the C-Mod dataset. These simple feature engineering methods give an alternative approach for improving models trained with tokamaks time series data.

Magnetic confinement devices such as Alcator C-Mod exhibit a variety of confinement regimes, including L-mode, H-mode, and intermediate modes like I-mode. The training dataset used in this work was compiled using features from DisruptionPy [4] and previous work [1] labeled transitions. An accurate and timely identification of regime transitions is especially important for the operation of next-generation devices like SPARC.

While previous efforts have assumed independent time observations models [1] or deep learning approaches for time series [3], the field of time series classification remains an active area of research with many alternative approaches (like advanced feature engineering) that we aim to explore in this work.

Supported by ENI.

Supported in part by Commonwealth Fusion Systems.

Supported by the DOE FES under Award DE-SC0024368.