Predicting Solar Flares Using a Long Short-Term Memory Network

Hao Liu; Chang Liu; Jason T. L. Wang; Haimin Wang

Source-linked AI summary

Predicting Solar Flares Using a Long Short-Term Memory Network

Hao Liu, Chang Liu, Jason T. L. Wang, Haimin Wang

arXiv:1905.07095v1May 17, 2019astro-ph.SRcs.LG

TL;DR

Solar-flare forecasting asks whether an active region will produce a gamma-class flare within 24 hours, a task motivated by the environmental effects of powerful flares. The paper models active-region observations as time series with LSTMs using magnetic and flare-history features, and reports better performance than related machine-learning methods while finding that reduced feature sets can outperform all 40 features.

Problem
The paper addresses prediction of ≥M5.0, ≥M, and ≥C flares from an active region within the next 24 hours, extending prior work that did not model active-region samples as time series with LSTMs.
Method
Three separate LSTM models use sequences of active-region samples containing magnetic and flare-history features, with an attention layer emphasizing relevant time steps.
Results
The LSTM approach outperforms related machine-learning methods and its ablations, while using 14–22 important features performs better than using all 40 features.
Takeaways & Limitations
Modeling active-region observations as time series provides the paper’s first reported application of LSTMs to solar flare prediction.
Takeaways & Limitations
Active regions are processed separately rather than combined into a single time series.

Abstract

from arXiv · show

We present a long short-term memory (LSTM) network for predicting whether an active region (AR) would produce a gamma-class flare within the next 24 hours. We consider three gamma classes, namely >=M5.0 class, >=M class, and >=C class, and build three LSTM models separately, each corresponding to a gamma class. Each LSTM model is used to make predictions of its corresponding gamma-class flares. The essence of our approach is to model data samples in an AR as time series and use LSTMs to capture temporal information of the data samples. Each data sample has 40 features including 25 magnetic parameters obtained from the Space-weather HMI Active Region Patches (SHARP) and related data products as well as 15 flare history parameters. We survey the flare events that occurred from 2010 May to 2018 May, using the GOES X-ray flare catalogs provided by the National Centers for Environmental Information (NCEI), and select flares with identified ARs in the NCEI flare catalogs. These flare events are used to build the labels (positive vs. negative) of the data samples. Experimental results show that (i) using only 14-22 most important features including both flare history and magnetic parameters can achieve better performance than using all the 40 features together; (ii) our LSTM network outperforms related machine learning methods in predicting the labels of the data samples. To our knowledge, this is the first time that LSTMs have been used for solar flare prediction.

1. INTRODUCTION

Solar flares motivate forecasting because powerful events and associated CMEs can severely affect the near-Earth environment, while their triggering mechanism remains incompletely understood. This paper introduces LSTMs for solar flare prediction by modeling active-region observations as time series.

Motivation: Powerful solar flares and associated CMEs can severely affect the near-Earth environment, motivating substantial forecasting and mitigation efforts.The triggering mechanism of solar flares is not fully understood.
Related work: Prior flare-prediction studies used magnetic, chromospheric, flare-history, and machine-learning information, including methods such as random forests, support vector machines, and neural networks.Related work includes Deep Flare Net, a deep neural network for flare prediction.
Approach: The paper uses SDO/HMI vector magnetic-field data and flaring history to predict whether an active region will produce a flare within 24 hours.The task is approached with a long short-term memory network.
Approach: LSTMs are suited to this task because observations within each active region form time series, allowing the network to learn order dependence between samples.The paper states that this is the first use of LSTMs for solar flare prediction.

2. DATA AND PREDICTIVE PARAMETERS

The study combines SHARP-derived magnetic measurements with flare-history features to represent active regions for flare prediction. It constructs samples from GOES/NCEI events and defines time-decay and count-based historical predictors.

Data collection: Flares from May 2010 through May 2018 were gathered from NCEI GOES catalogs, retaining events with identified active regions and querying SHARP-related data from JSOC.The resulting database contains 4,203 B-class, 6,768 C-class, 704 M-class, and 49 X-class flares.
Predictive parameters: The dataset combines 25 SDO/HMI magnetic parameters with 15 flare-history features, yielding 40 predictive features per sample.The magnetic parameters characterize active-region magnetic-field properties, while the history features describe prior flare activity.
Flare history: Six flare-history features encode time-decayed prior activity, with class-specific values based on earlier flares in the same active region.The decay constant τ is set to 12, and Figure 1 illustrates the M-class calculation.
Flare history: The remaining nine flare-history features count prior B-, C-, M-, and X-class flares overall and during the preceding 24 hours, with an additional maximum-flare measure.These features are described as historical predictors for each sample in an active region.
Preprocessing: Feature values are normalized to account for differences in units and scales before modeling.The physical features use mean and standard-deviation normalization, while the supplied passages also describe feature extrema for scaling.

3. METHODOLOGY

The method formulates flare prediction as three separate 24-hour binary classification tasks and models each active region’s observations as a time series. An attention-based LSTM processes ten-sample sequences built from magnetic, flare-history, and temporally structured observations, producing positive or negative predictions.

3.2. Prediction Method: Training, validation, and testing use 2010–2013, 2014, and 2015–2018 data respectively, with disjoint training and testing active regions.Observations outside ±70° of central meridian or with incomplete features are ignored.
3.1. Prediction Task: Three separate models predict whether an active region will produce a ≥M5.0, ≥M, or ≥C flare within the next 24 hours.Each model corresponds to one flare class.
3.1. Prediction Task: Positive samples are the 24 hourly observations preceding a target flare, while the flare-occurrence hour is excluded from the dataset.Other observations are assigned to the negative class.
3.2. Prediction Method: LSTM units update a cell state through forget, input, and output gates, allowing the model to retain and use information across ordered observations.The cell state combines the previous cell state with a candidate cell state.
3.2. Prediction Method: Attention weights compare each time-step state with the final state to form a context vector that is combined with the last hidden state.The resulting attention vector is passed to the fully connected layers.
3.1. Prediction Task: Each training example uses m consecutive active-region samples, with the sequence label determined by the final sample’s positive or negative class.During testing, the model receives the current sample and its preceding m−1 samples.
3.2. Prediction Method: The procedure handles class imbalance with weighted cross entropy and processes each active region separately using zero-padded sequences of ten samples.The positive minority class receives greater weight during optimization.
3.2. Prediction Method: The network combines an LSTM layer, an attention layer, two fully connected layers, and a two-neuron softmax output layer.The architecture uses 200 neurons in the first fully connected layer, 500 in the second, and m LSTM and attention units.

4. RESULTS

The experiments evaluate LSTM architectures, feature subsets, and comparisons with related methods for predicting flares within 24 hours. LSTM performs significantly better than RF as a binary classifier across all three flare classes, although RF performs better for probabilistic forecasting in the reported comparisons.

4.2. Model Evaluation: The evaluation compares four LSTM architectures and five related machine-learning methods using flare prediction within 24 hours.The study also uses cross-validation with means and standard deviations calculated over 100 runs.
4.3. Feature Assessment: Sixteen features are consistently ranked in the top 20 across all three models, including ten magnetic parameters and several flare-history features.The listed features include magnetic quantities such as TOTUSJH and USFLUX, alongside Cdec, Chis, Chis1d, Edec, Mhis, and Xmax1d.
4.3. Feature Assessment: Using the top 20, 22, and 14 features gives the highest mean cumulative TSS for ≥M5.0, ≥M, and ≥C models, respectively, rather than using all 40 features.The selected subsets are approximately 14–22 of the 40 available features.
4.4. Comparison between RF and LSTM: LSTM significantly outperforms RF in BACC, HSS, and TSS across all three flare classes when both methods are used as binary classifiers.The comparison uses cross-validation and reports p < 0.05 for every flare class and metric.
4.4. Comparison between RF and LSTM: For ≥M5.0 flares, LSTM reaches TSS=0.881, BACC=0.940±0.007, HSS=0.084±0.015, and AUC=0.984±0.003, exceeding RF on these metrics.RF achieves TSS=0.812, BACC=0.906±0.014, HSS=0.042±0.002, and AUC=0.948±0.011 in the same comparison.
4.4. Comparison between RF and LSTM: RF is better than LSTM for probabilistic forecasting on the reported ≥M5.0, ≥M, and ≥C comparisons based on BS and BSS.The ≥M and ≥C passages explicitly state that RF is the better probabilistic forecasting model, while the ≥M5.0 comparison reports better BS and BSS for RF.

5. DISCUSSION AND CONCLUSIONS

The study applies separate LSTM models to three flare classes, treating active-region samples as time series and combining magnetic and flare-history features. Results support selective feature use and show that LSTM and random forests are preferable for different forecasting objectives.

Experimental setting: The framework uses three class-specific LSTM models and 40 features drawn from SHARP magnetic data, related products, and flare history.The dataset covers May 2010 through May 2018, with disjoint training and testing periods so predictions target unseen active regions.
Methodological contribution: The paper’s central methodological contribution is modeling active-region samples as time series so LSTMs can capture temporal dependencies for solar-flare prediction.The authors identify this as the first use of LSTMs for solar-flare prediction.
Feature analysis: Using 14–22 important magnetic and flare-history features performs better than using all 40 features, with C-class flare history contributing most among history parameters.Several magnetic parameters and flare-history features are consistently predictive across the three LSTM models.
Main results: LSTM outperforms MLP, JN, SVM, Deep Flare Net, and its ablations, while RF remains comparable overall.The proposed architecture benefits from adding attention and fully connected layers to LSTM units.
Main results: RF is recommended for probabilistic flare estimates, whereas LSTM is significantly better for firm yes-or-no predictions within 24 hours.For ≥C-class prediction, LSTM has higher TSS, BACC, HSS, and AUC, while RF has better BS and BSS.
Conclusion: The authors conclude that flare history complements SDO/HMI magnetic parameters and plan to incorporate solar images in future deep-learning flare prediction.They also identify filament eruptions and CMEs as possible future prediction targets.