Solar Datasets Research Articles

Automated High-Precision Recognition of Solar Filaments Based on an Improved U2-Net

Solar filaments are a significant solar activity phenomenon, typically observed in full-disk solar observations in the H-alpha band. They are closely associated with the magnetic fields of solar active regions, solar flare eruptions, and coronal mass ejections. With the increasing volume of observational data, the automated high-precision recognition of solar filaments using deep learning is crucial. In this study, we processed full-disk H-alpha solar images captured by the Chinese H-alpha Solar Explorer in 2023 to generate labels for solar filaments. The preprocessing steps included limb-darkening removal, grayscale transformation, K-means clustering, particle erosion, multiple closing operations, and hole filling. The dataset containing solar filament labels is constructed for deep learning. We developed the Attention U2-Net neural network for deep learning on the solar dataset by introducing an attention mechanism into U2-Net. In the results, Attention U2-Net achieved an average Accuracy of 0.9987, an average Precision of 0.8221, an average Recall of 0.8469, an average IoU of 0.7139, and an average F1-score of 0.8323 on the solar filament test set, showing significant improvements compared to other U-net variants.

Improving Earth-like planet detection in radial velocity using deep learning

Context. Many novel methods have been proposed to mitigate stellar activity for exoplanet detection as the presence of stellar activity in radial velocity (RV) measurements is the current major limitation. Unlike traditional methods that model stellar activity in the RV domain, more methods are moving in the direction of disentangling stellar activity at the spectral level. As deep neural networks have already been proven to be one of the most effective tools in data mining, in this work, we explore their potential in the context of Earth-like planet detection in RV measurements. Aims. The goal of this paper is to present a novel convolutional neural network-based algorithm that efficiently models stellar activity signals at the spectral level, enhancing the detection of Earth-like planets. Methods. Based on the idea that the presence of planets can only produce a Doppler shift at the spectral level while the presence of stellar activity can introduce a variation in the profile of spectral lines (asymmetry and depth change), we trained a convolutional neural network to build the correlation between the change in the spectral line profile and the corresponding RV, full width at half maximum (FWHM) and bisector span (BIS) values derived from the classical cross-correlation function. Results. This algorithm has been tested on three intensively observed stars: Alpha Centauri B (HD 128621), Tau ceti (HD 10700), and the Sun. By injecting simulated planetary signals at the spectral level, we demonstrate that our machine learning algorithm can achieve, for HD 128621 and HD 10700, a detection threshold of 0.5 m s−1 in semi-amplitude for planets with periods ranging from 10 to 300 days. This threshold would correspond to the detection of a ~4 M⊕ in the habitable zone of those stars. On the HARPS-N solar dataset, thanks to significantly more data, our algorithm is even more efficient at mitigating stellar activity signals and can reach a threshold of 0.2 m s−1, which would correspond to a 2.2 M⊕ planet on the orbit of the Earth. Conclusions. To the best of our knowledge, it is the first time that such low detection thresholds are reported for the Sun, but also for other stars, and therefore this highlights the efficiency of our convolutional neural network-based algorithm at mitigating stellar activity in RV measurements.

Short-term Classification of Strong Solar Energetic Particle Events Using Multivariate Time-series Classifiers

Solar energetic particle (SEP) events are one of the most crucial aspects of space weather that require continuous monitoring and forecasting. Their prediction depends on various factors, including source eruptions. In the present work, we use the Geostationary Solar Energetic Particle data set covering solar cycles 22, 23, and 24. We develop a framework using time-series-based machine-learning (ML) models with the aim of developing robust short-term forecasts by classifying SEP events. For this purpose, we introduce an ensemble learning approach that merges the results from univariate time series of three proton channels (E ≥10, 50, and 100 MeV) and the long-band X-ray flux (1–8 Å) channel from the Geostationary Operational Environmental Satellite missions and analyze their performance. We consider three models, namely, time series forest, supervised time series forest (STSF), and Bag-of-Symbolic Fourier Approximation Symbols. Our study also focuses on understanding and developing confidence in the predictive capabilities of our models. Therefore, we utilize multiple evaluation techniques and metrics. Based on that, we find STSF to perform well in all scenarios. The summary of metrics for the STSF model is as follows: the area under the ROC curve = 0.981, F 1-score = 0.960, true skill statistics = 0.919, Heidke skill score = 0.920, Gilbert skill score = 0.852, and Matthew’s correlation coefficient = 0.920. The Brier score loss of the STSF model is 0.077. This work lays the foundation for building near-real-time short-term SEP event predictions using robust ML methods.

Geospatial optimisation of wind and solar energy generation: A framework for optimised resource allocation

Wind and solar power generation profiles exhibit a high degree of power output variability and uncertainty compared to conventional fossil fuel generation plants. However, it is generally accepted that this variability and uncertainty can be mitigated by geographic siting and dispersion of the renewable energy plants, particularly with the view to leverage aggregation effects. In this context, this paper proposes a framework for the geospatial optimisation of wind and solar PV generation for scenarios with a high penetration of wind and solar renewable energy. The proposed optimisation framework represents a probabilistic risk-based approach that seeks to minimise the number of events where high residual load values require ancillary service interventions to maintain power balance. In this approach, the wind and solar resources are categorised in terms of the statistical properties of the associated spatiotemporal wind and solar power profiles for a given set of daily and seasonal Time of Use periods. It is thereby recognised that the resource characteristics and grid impact of wind and solar generation profiles can be interpreted with reference to the daily and seasonal cycles exhibited by the demand profiles, wherein some Time of Use periods are more important than others. The proposed framework is implemented for a number of baseline case studies and optimisation case studies, using wind and solar resource data sets covering South Africa. Overall, it is concluded that this approach can leverage geospatial site optimisation to substantially reduce the risks associated with high residual load events. Furthermore, the proposed framework is highly flexible in that the formulation of the minimum and maximum allocation constraints allows application for real-world scenarios whereby stringent capacity allocation constraints can be applied.

An Experimental Study on EUV‐To‐Magnetogram Image Translation Using Conditional Generative Adversarial Networks

AbstractDeep generative models have recently become popular in heliophysics for their capacity to fill in gaps in solar observational data sets, thereby helping mitigating the data scarcity issue faced in space weather forecasting. A particular type of deep generative models, called conditional Generative Adversarial Networks (cGAN), has been used since a few years in the context of image‐to‐image (I2I) translation on solar observations. These algorithms have however hyperparameters whose values might influence the quality of the synthetic image. In this work, we use magnetograms produced by the Helioseismic and Magnetic Imager (HMI) on board the Solar Dynamics Observatory (SDO) and EUV images from the Atmospheric Imaging Assembly (AIA) for the problem of generating Artificial Intelligence (AI) synthetic magnetograms from multiple SDO/AIA channels using cGAN, and more precisely the Pix2PixCC algorithm. We perform a systematic study of the most important hyperparameters to investigate which hyperparameter might generate magnetograms of highest quality with respect to the Structural Similarity Index. We propose a structured way to perform training with various hyperparameter values, and provide diagnostic and visualization tools of the generated versus targeted image. Our results shows that when using a larger number of filters in the convolution blocks of the cGAN, the fine details in the generated magnetogram are better reconstructed. Adding several input channels besides the 304 Å channel does not improve the quality of generated magnetogram, but the hyperparameters controlling the relative importance of the different loss functions in the optimization process have an influence on the quality of the results.

Staring at the Sun with the Keck Planet Finder: An Autonomous Solar Calibrator for High Signal-to-noise Sun-as-a-star Spectra

Extreme precision radial velocity (EPRV) measurements contend with internal noise (instrumental systematics) and external noise (intrinsic stellar variability) on the road to 10 cm s−1 “exo-Earth” sensitivity. Both of these noise sources are well-probed using “Sun-as-a-star” RVs and cross-instrument comparisons. We built the Solar Calibrator (SoCal), an autonomous system that feeds stable, disk-integrated sunlight to the recently commissioned Keck Planet Finder (KPF) at the W. M. Keck Observatory. With SoCal, KPF acquires signal-to-noise ratio (S/N) ∼ 1200, R = 98,000 optical (445–870 nm) spectra of the Sun in 5 s exposures at unprecedented cadence for an EPRV facility using KPF’s fast readout mode (<16 s between exposures). Daily autonomous operation is achieved by defining an operations loop using state machine logic. Data affected by clouds are automatically flagged using a reliable quality control metric derived from simultaneous irradiance measurements. Comparing solar data across the growing global network of EPRV spectrographs with solar feeds will allow EPRV teams to disentangle internal and external noise sources and benchmark spectrograph performance. To facilitate this, all SoCal data products are immediately available to the public on the Keck Observatory Archive. We compared SoCal RVs to contemporaneous RVs from NEID, the only other immediately public EPRV solar data set. We find agreement at the 30–40 cm s−1 level on timescales of several hours, which is comparable to the combined photon-limited precision. Data from SoCal were also used to assess a detector problem and wavelength calibration inaccuracies associated with KPF during early operations. Long-term SoCal operations will collect upwards of 1000 solar spectra per six-hour day using KPF’s fast readout mode, enabling stellar activity studies at high S/N on our nearest solar-type star.

Solar Radio Burst Detection Based on the MobileViT-SSDLite Lightweight Model

Real-time detection of solar radio bursts is crucial in solar physics research and space weather forecasting. However, current research on the automatic detection of solar radio bursts is limited to identifying the presence or absence of solar radio bursts or recognizing only a single type of burst, such as type II or III. Furthermore, existing methods cannot learn spectral and temporal features and often suffer from the drawbacks of large network models, resulting in slow speeds. This paper proposes an automatic recognition and localization method based on a lightweight object detection model for solar radio burst events. We collected observation data from e-CALLISTO and established a data set containing type II, III, IV, and V solar radio bursts. To address the real-time requirements of practical applications and consider the temporal and frequency domain information of spectrogram images, we improved a vision transformer with a self-attention mechanism and adopted a lightweight model for detection. The experimental results demonstrate that our proposed method achieves an average precision at a 50% intersection-over-union threshold of 78.2% and a recall rate of 92% on the established solar radio burst data set. Additionally, the model operates at a detection speed of 54.8 frames s–1, where a frame refers to a spectral image with a duration of 15 minutes, enabling efficient automated detection and localization of type II, III, IV, and V solar radio bursts.

Eagle arithmetic optimization algorithm for renewable energy-based load frequency stabilization of power systems

Power systems' efficient management and planning are crucial in renewable energy-based systems. As the global electricity demand continues to rise, there is a growing need for alternative energy sources such as solar, wind, and hydropower. Consequently, numerous research studies have focused on maintaining load balancing within the renewable energy system and improving the forecasting of renewable energy resources. This paper presents the Eagle Arithmetic Optimization Algorithm (EAOA) as a novel approach to address these challenges. By utilizing a fuzzy-based dragonfly optimization algorithm (fuzzy-DFOA), the proposed method enhances the accuracy of load-balancing analysis in renewable energy resources. Through its innovative techniques, the EAOA demonstrates its potential to significantly improve the efficiency and effectiveness of managing renewable energy systems, paving the way for a more sustainable and reliable power grid. The accuracy rate of both wind and solar datasets is given. For the wind dataset, our proposed work got 92.63%, SVR got 75.89%, CNN got 87.54%, and QODA got 83.16%. For the solar dataset presented work of fuzzy-based DFOA got 92.59%, SVR got 69.16%, CNN got 86.25%, and QODA got 82.37%.

Challenges in the Detection and Attribution of Northern Hemisphere Surface Temperature Trends Since 1850

Since 2007, the Intergovernmental Panel on Climate Change (IPCC) has heavily relied on the comparison between global climate model hindcasts and global surface temperature (ST) estimates for concluding that post-1950s global warming is mostly human-caused. In Connolly et al., we cautioned that this approach to the detection and attribution of climate change was highly dependent on the choice of Total Solar Irradiance (TSI) and ST data sets. We compiled 16 TSI and five ST data sets and found by altering the choice of TSI or ST, one could (prematurely) conclude anything from the warming being “mostly human-caused” to “mostly natural.” Richardson and Benestad suggested our analysis was “erroneous” and “flawed” because we did not use a multilinear regression. They argued that applying a multilinear regression to one of the five ST series re-affirmed the IPCC’s attribution statement. They also objected that many of the published TSI data sets were out-of-date. However, here we show that when applying multilinear regression analysis to an expanded and updated data set of 27 TSI series, the original conclusions of Connolly et al. are confirmed for all five ST data sets. Therefore, it is still unclear whether the observed warming is mostly human-caused, mostly natural or some combination of both.

Solar panel defect detection design based on YOLO v5 algorithm

Defects of solar panels can easily cause electrical accidents. The YOLO v5 algorithm is improved to make up for the low detection efficiency of the traditional defect detection methods. Firstly, it is improved on the basis of coordinate attention to obtain a LCA attention mechanism with a larger target range, which can enhance the sensing range of target features in addition to fully capturing feature information; secondly, the weighted bidirectional feature pyramid is used to balance the feature information with excessive pixel differences by assigning different weights, which is more conducive to multi-scale Fast fusion of features; finally, the usual coupled head of YOLO series is replaced with decoupled head, so that the task branch can be performed more accurately and the detection accuracy can be improved. The results of comparative experiments on the solar panel defect detection data set show that after the improvement of the algorithm, the overall precision is increased by 1.5%, the recall rate is increased by 2.4%, and the mAP is up to 95.5%, which is 2.5% higher than that before the improvement. It can more accurately determine whether there are defects, standardize the quality of solar panels, and ensure electrical safety.

Unsigned magnetic flux proxy from solar optical intensity spectra

ABSTRACT The photospheric unsigned magnetic flux has been shown to be highly correlated with radial velocity (RV) variations caused by solar surface activity. This activity indicator is therefore a prime candidate to unlock the potential of RV surveys to discover Earth twins orbiting Sun-like stars. We show for the first time how a precise proxy of the unsigned magnetic flux (ΔαB2) can be obtained from Sun-as-a-star intensity spectra by harnessing the magnetic information contained in over 4000 absorption lines in the wavelength range from 380 to 690 nm. This novel activity proxy can thus be obtained from the same spectra from which RVs are routinely extracted. We derived ΔαB2 from 500 randomly selected spectra from the HARPS-N public solar data set, which spans from 2015 to 2018. We compared our estimates with the unsigned magnetic flux values from the Solar Dynamics Observatory (SDO) finding excellent agreement (median absolute deviation: 4.9 per cent). The extracted indicator ΔαB2 correlates with SDO’s unsigned magnetic flux estimates on the solar rotational time-scale (Pearson correlation coefficient 0.67) and on the 3-yr time-scale of our data set (correlation coefficient 0.91). We find correlations of ΔαB2 with the HARPS-N solar RV variations of 0.49 on the rotational time-scale and 0.78 on the 3-yr time-scale. The Pearson correlation of ΔαB2 with the RVs is found to be greater than the correlation of the classical activity indicators with the RVs. For solar-type stars, ΔαB2 therefore represents the best simultaneous activity proxy known to date.

Short Term Solar Irradiation Prediction Framework Based on EEMD-GA-LSTM Method

Accurate short term solar irradiation forecasting is necessary for smart grid stability and to manage bilateral contract negotiations between suppliers and customers. Traditional machine learning methods are unable to acquire and rectify nonlinear characteristics from solar dataset, which not only complicates model construction but also affect prediction accuracy. To address these issues, a deep learning based architecture with predictive analysis strategy is developed in this manuscript. In the first stage, the original solar irradiation sequences are divided into many intrinsic mode functions to generate a prospective feature set using a sophisticated signal decomposition technique. After that, an iteration method is used to generate a prospective range of frequency related to deep learning model. This method is created by linked algorithm using the GA and deep learning network. The findings by the proposed model employing sequences obtained by the preprocessing methodology considerable improve prediction accuracy as comparison to conventional models. In contrast, when confronted with a high resolution dataset derived from big data set, the chosen dataset may not only conduct a huge data reduction, but also enhances forecasting accuracy up to 22.74 percent over a variety of evaluation metrics. As a result, the proposed method might be used to predict short-term solar irradiation with greater accuracy using a solar dataset.

Solar Power Prediction Using Dual Stream CNN-LSTM Architecture.

The integration of solar energy with a power system brings great economic and environmental benefits. However, the high penetration of solar power is challenging due to the operation and planning of the existing power system owing to the intermittence and randomicity of solar power generation. Achieving accurate predictions for power generation is important to provide high-quality electric energy for end-users. Therefore, in this paper, we introduce a deep learning-based dual-stream convolutional neural network (CNN) and long short-term nemory (LSTM) network followed by a self-attention mechanism network (DSCLANet). Here, CNN is used to learn spatial patterns and LSTM is incorporated for temporal feature extraction. The output spatial and temporal feature vectors are then fused, followed by a self-attention mechanism to select optimal features for further processing. Finally, fully connected layers are incorporated for short-term solar power prediction. The performance of DSCLANet is evaluated on DKASC Alice Spring solar datasets, and it reduces the error rate up to 0.0136 MSE, 0.0304 MAE, and 0.0458 RMSE compared to recent state-of-the-art methods.

Fast Quasi-Static Time-Series Simulation for Accurate PV Inverter Semiconductor Fatigue Analysis with a Long-Term Solar Profile

Power system simulations with long-term data typically have large time steps, varying from one second to a few minutes. However, for PV inverter semiconductors in grid-connected applications, the minimum thermal stress cycle occurs over the fundamental grid frequency (50 or 60 Hz). This requires the time step of the fatigue simulation to be approximately 100 μs. This small time step requires long computation times to process yearly power production profiles. In this paper, we propose a fast fatigue simulation for inverter semiconductors using the quasi-static time-series simulation concept. The proposed simulation calculates the steady state of the semiconductor junction temperature using a fast Fourier transform. The small thermal cycling during a switching period and even over the fundamental waveform is disregarded to further accelerate the simulation speed. The resulting time step of the fatigue simulation is 15 min, which is consistent with the solar dataset. The error of the proposed simulation is 0.16% compared to the fatigue simulation results using the complete thermal stress profile. The error of the proposed method is significantly less than the conventional averaged thermal profile. A PV inverter that responds to a transactive energy system is simulated to demonstrate the use of the proposed fatigue simulation. The proposed simulation has the potential to cosimulate with system-level simulation tools that also adopt the quasi-static time-series concept.

Erroneous use of Statistics behind Claims of a Major Solar Role in Recent Warming

In a study that attempted to relate solar and human activity to Earth's recent temperature change, Connolly et al. committed a basic error in the choice of statistical methods and so overreported the effect of the Sun. A major theme of their study was that there are many data sets of past solar activity, and some of these allegedly provide statistical evidence of “most of the recent global warming being due to changes in solar activity.” We avoid methods that are known to give inaccurate results and show that for 1970–2005 Northern Hemisphere land the corrected solar attribution fraction is −7% to +5%, compared with values of up to 64% reported in Connolly et al. Their higher values are entirely due to mistaken application of statistics. Unfortunately, we cannot test truly “recent” global warming since most of their solar data sets end before 2015, and two finish in the 1990s, but all tested post-1970 periods show similarly small solar contributions. The solar-climate linkage is an area of fascinating and ongoing research with rigorous technical discussion. We argue that instead of repeating errors, they should be acknowledged and corrected so that the debate can focus on areas of legitimate scientific uncertainty.

Constraining super-light sterile neutrinos at Borexino and KamLAND

The presence of a super-light sterile neutrino can lead to a dip in the survival probability of solar neutrinos, and explain the suppression of the upturn in the low energy solar neutrino data. In this work, we systematically study the survival probabilities in the 3+1 framework by taking into account of the non-adiabatic transitions and the coherence effect. We obtain an analytic equation that can predict the position of the dip. We also place constraints on the parameter space of sterile neutrinos by using the latest Borexino and KamLAND data. We find that the low and high energy neutrino data at Borexino are sensitive to different regions in the sterile neutrino parameter space. In the case with only θ01 being nonzero, the 8B data sets the strongest bounds at Delta {m}_{01}^2 ≈ (1.1 ∼ 2.2) Delta {m}_{21}^2 , while the low energy neutrino data is more sensitive to other mass-squared regions. The lowest bounds on Delta {m}_{01}^2 from the pp data can reach 10−12 eV2 because of the coherence effect. Also, due to the presence of non-adiabatic transitions, the bounds in the range of 10−9 eV2 ≲ Delta {m}_{01}^2 ≲ 10−5 eV2 become weaker as Delta {m}_{01}^2 or sin2 2θ01 decreases. We also find that in the case with only θ02 or θ03 being nonzero, the low energy solar neutrino data set similar but weaker bounds as compared to the case with only θ01 being nonzero. However, the bounds from the high energy solar data and the KamLAND data are largely affected by the sterile mixing angles.

Country-wide solar power load profile for Germany 2015 to 2019: The impact of system curtailments on prediction models

A large (43824 h) country-wide solar power load profile (LP), solar irradiance and meteorological dataset (ten variables) for Germany covering years 2015 to 2019 is compiled and forecast with eight machine learning and deep learning (ML/DL) algorithms. Analysis reveals that system curtailments are likely responsible for some outlying predictions. The adaptive boosting (ADA) and random forest (RF) algorithms outperform convolutional neural networks and other algorithms in supervised prediction and forecasting tasks with the dataset. Once tuned with 2015 to 2018 data, ADA and RF forecast 2019 hourly data with root mean squared error (RMSE) of < 0.03. Similar forecasting accuracy is achieved using smaller datasets of three months of historical data from 2015 to 2018 to forecast hourly LP in each month of 2019. Forecasts for July 2019 are associated with the highest errors (RMSE = 0.41). The transparent open box (TOB) algorithm, because it can reveal details of data matching contributions to each of its forecasts, is used to data mine the July 2019 forecasts and conduct in-depth outlier analysis. It reveals nine extreme outliers that each overestimate LP based on irradiance and meteorological inputs. Analysis suggests that substantial overestimates leading to high RMSE for July 2019 are likely due to system curtailments. It is beneficial to combine less transparent (e.g., ADA and RF) with more transparent (e.g. TOB) ML algorithms to accurately forecast and data mine large solar power data sets. It also indicates that solar LP forecasts cannot always rely on irradiance and meteorological variables in isolation. There is a need to be mindful of system constraints and market conditions when predicting LP on a country-wide basis.

Short Term Solar Irradiation Forecasting using CEEMDAN Decomposition Based BiLSTM Model Optimized by Genetic Algorithm Approach

An accurate short-term solar irradiation forecasting is requiredregarding smart grid stability and to conduct bilateral contract negotiations between suppliers and customers. Traditional machine learning models are unable to acquire and to rectify nonlinear properties from solar datasets, which not only complicate model formation but also lower prediction accuracy. The present research paper develops a deep learningbased architecture with a predictive analytic technique to address these difficulties. Using a sophisticated signal decomposition technique, the original solar irradiation sequences are decomposed into multiple intrinsic mode functions to build a prospective feature set. Then, using an iteration strategy, a potential range of frequency associated to the deep learning model is generated. This method is developed utilizing a linked algorithm and a deep learning network. In comparison with conventional models, the suggested model utilizes sequences generated through preprocessing methods, significantly improving prediction accuracywhen confronted with a high resolution dataset created from a big dataset.On the other hand, the chosen dataset not only performs a massive data reduction, but also improves forecasting accuracy by up to 20.74 percent across a range of evaluation measures. The proposed model achieves lowest annual average RMSE (1.45W/m2), MAPE (2.23%) and MAE (1.34W/m2) among the other developed models for 1-hr ahead solar GHI, respectively, whereas forecast-skill obtained by the proposed model is 59% with respect to benchmark model. As a result, the proposed method might be used to predict short-term solar irradiation with greater accuracy using a solar dataset

Comprehensive assessment, review, and comparison of AI models for solar irradiance prediction based on different time/estimation intervals

Solar energy-based technologies have developed rapidly in recent years, however, the inability to appropriately estimate solar energy resources is still a major drawback for these technologies. In this study, eight different artificial intelligence (AI) models namely; convolutional neural network (CNN), artificial neural network (ANN), long short-term memory recurrent model (LSTM), eXtreme gradient boost algorithm (XG Boost), multiple linear regression (MLR), polynomial regression (PLR), decision tree regression (DTR), and random forest regression (RFR) are designed and compared for solar irradiance prediction. Additionally, two hybrid deep neural network models (ANN-CNN and CNN-LSTM-ANN) are developed in this study for the same task. This study is novel as each of the AI models developed was used to estimate solar irradiance considering different timesteps (hourly, every minute, and daily average). Also, different solar irradiance datasets (from six countries in Africa) measured with various instruments were used to train/test the AI models. With the aim to check if there is a universal AI model for solar irradiance estimation in developing countries, the results of this study show that various AI models are suitable for different solar irradiance estimation tasks. However, XG boost has a consistently high performance for all the case studies and is the best model for 10 of the 13 case studies considered in this paper. The result of this study also shows that the prediction of hourly solar irradiance is more accurate for the models when compared to daily average and minutes timestep. The specific performance of each model for all the case studies is explicated in the paper.

Decreasing False-alarm Rates in CNN-based Solar Flare Prediction Using SDO/HMI Data

A hybrid two-stage machine-learning architecture that addresses the problem of excessive false positives (false alarms) in solar flare prediction systems is investigated. The first stage is a convolutional neural network (CNN) model based on the VGG-16 architecture that extracts features from a temporal stack of consecutive Solar Dynamics Observatory Helioseismic and Magnetic Imager magnetogram images to produce a flaring probability. The probability of flaring is added to a feature vector derived from the magnetograms to train an extremely randomized trees (ERT) model in the second stage to produce a binary deterministic prediction (flare/no-flare) in a 12 hr forecast window. To tune the hyperparameters of the architecture, a new evaluation metric is introduced: the “scaled True Skill Statistic.” It specifically addresses the large discrepancy between the true positive rate and the false positive rate in the highly unbalanced solar flare event training data sets. Through hyperparameter tuning to maximize this new metric, our two-stage architecture drastically reduces false positives by ≈48% without significantly affecting the true positives (reduction by ≈12%), when compared with predictions from the first-stage CNN alone. This, in turn, improves various traditional binary classification metrics sensitive to false positives, such as the precision, F1, and the Heidke Skill Score. The end result is a more robust 12 hr flare prediction system that could be combined with current operational flare-forecasting methods. Additionally, using the ERT-based feature-ranking mechanism, we show that the CNN output probability is highly ranked in terms of flare prediction relevance.