SH21A-4092:
Solar Flare Prediction Using SDO/HMI Vector Magnetic Field Data with a Machine-Learning Algorithm

Tuesday, 16 December 2014
Monica Bobra and Sebastien P Couvidat, Stanford University, Stanford, CA, United States
Abstract:
We attempt to forecast M-and X-class solar flares using a machine-learning algorithm, called Support Vector Machine (SVM), and four years of data from the Solar Dynamics Observatory's Helioseismic and Magnetic Imager, the first instrument to continuously map the full-disk photospheric vector magnetic field from space (Schou et al., 2012). Most flare forecasting efforts described in the literature use either line-of-sight magnetograms or a relatively small number of ground-based vector magnetograms. This is the first time such a large dataset of vector magnetograms has been used to forecast solar flares.

We build a catalog of flaring and non-flaring active regions sampled from a database of 2,071 active regions, comprised of 1.5 million active region patches of vector magnetic field data, and characterize each active region by 25 parameters --- which include the flux, energy, shear, current, helicity, gradient, geometry, and Lorentz force. We then train and test the machine-learning algorithm. Finally, we estimate the performance of this algorithm using forecast verification metrics with an emphasis on the true skill statistic (TSS). Bloomfield et al. (2012) suggest the use of the TSS as it is not sensitive to the class imbalance problem. Indeed, there are many more non-flaring active regions in a given time interval than flaring ones: this class imbalance distorts many performance metrics and renders comparison between various studies somewhat unreliable.

We obtain relatively high TSS scores and overall predictive abilities. We surmise that this is partly due to fine-tuning the SVM for this purpose and also to an advantageous set of features that can only be calculated from vector magnetic field data. We also apply a feature selection algorithm to determine which of our 25 features are useful for discriminating between flaring and non-flaring active regions and conclude that only a handful are needed for good predictive abilities.