Reliable Probability Forecast of Solar Flares: Deep Flare Net-Reliable (DeFN-R)

TL;DR

This paper introduces DeFN-R, a deep neural network model that provides reliable probabilistic forecasts of solar flares within 24 hours, improving prediction reliability and offering adjustable probability thresholds for operational use.

Contribution

The paper presents DeFN-R, a novel deep learning model that enhances the reliability of probabilistic solar flare forecasts using a large dataset and optimized training for Brier skill score.

Findings

Achieved BSS = 0.41 for C-class flares

Achieved BSS = 0.30 for M-class flares

Improved forecast reliability while maintaining ROC performance

Abstract

We developed a reliable probabilistic solar flare forecasting model using a deep neural network, named Deep Flare Net-Reliable (DeFN-R). The model can predict the maximum classes of flares that occur in the following 24 h after observing images, along with the event occurrence probability. We detected active regions from 3x10^5 solar images taken during 2010-2015 by Solar Dynamic Observatory and extracted 79 features for each region, which we annotated with flare occurrence labels of X-, M-, and C-classes. The extracted features are the same as used by Nishizuka et al. (2018); for example, line-of-sight/vector magnetograms in the photosphere, brightening in the corona, and the X-ray emissivity 1 and 2 h before an image. We adopted a chronological split of the database into two for training and testing in an operational setting: the dataset in 2010-2014 for training and the one in 2015 for testing. DeFN-R is composed of multilayer perceptrons formed by batch normalizations and skip connections. By tuning optimization methods, DeFN-R was trained to optimize the Brier skill score (BSS). As a result, we achieved BSS = 0.41 for >=C-class flare predictions and 0.30 for >=M-class flare predictions by improving the reliability diagram while keeping the relative operating characteristic curve almost the same. Note that DeFN is optimized for deterministic prediction, which is determined with a normalized threshold of 50%. On the other hand, DeFN-R is optimized for a probability forecast based on the observation event rate, whose probability threshold can be selected according to users' purposes.