The S3EP-AC provides an All-Clear SEP event forecast, and is composed of a set of connected space weather event forecasting modules, working in different modes. The All-Clear SEP event forecast places emphasis on precise and sensitive prediction of non-flaring active regions (CBN; labeling as flare quiet active regions or ones hosting lower-class [A, B, C] flares) to identify periods where the occurrence of an SEP event is highly unlikely. In such cases, SEP events of >10 MeV proton flux are not likely to exceed 10 pfu. Later phases of development will implement the SEP Watch, where conditions are likely to lead to an SEP triggering event, and an SEP Warning, where a triggering event has occurred and an SEP event is likely imminent.
The connected modules of the system are constructed as ensembles of different learning algorithms, producing binary, probabilistic, and regression-based prediction models for flare, eruptive-flare, and CME speed predictors. All predictions are designated for and deployed on individual active regions. All reports are aggregated for a final probabilistic ‘All-Clear’ output based on a set of user-defined thresholds. The All-Clear SEP predictor utilizes NRT HARP data multi variate time series (MVTS) of AR metadata, with many of the metadata parameter calculation algorithms being the same as the SHARP keywords. All models are trained for a 12-hour observation window, meaning that a complete prediction result can be issued only after 12 hours worth of high-quality data has been collected. The prediction window (forecast validity period) for all the models is 24 hours with a zero latency (i.e., forecasts are effective immediately).
The predictive process follows three distinct paths. First, it determines the probability of a sizable flare (i.e., M+) occurrence within the next 24 hours. This path uses three base learners (i.e., SOHO-FP, DSDO-FP, NSDO-FP) and a meta learner (which uses the output of base learners), described in Ji et al. (2020), where each base learner is a multivariate time series classifier (MTSC) based on the Time Series Forest algorithm (Deng et al., 2013}. The second path predicts the probability of an eruptive flare occurrence within the next 24 hours. This path also uses the AR MVTS and issues probabilities for occurrence of eruptive (P(ER)) vs. non-eruptive (P(NE)) events. Note here that an eruption may originate from X-, M-, or C-class flares, but not all flares are eruptive. A- and B-class flares are not considered in this framework. The third path uses the outputs of base learners to predict the occurrence probability of X-, M-, C-class flares and Flare-Quiet (FQ) regions. This is a quaternary meta-learner with its outputs fed into a regressor seeking to project a CME speed.
As the prediction algorithms of this system are trained on individual active regions, the results of individual forecasts are then aggregated via (1) thresholded activation functions and (2) a multiplicative model which assumes conditional independence of the active regions. Namely, the full disk all clear probability is then calculated as the joint all-clear probability of active region all-clear outputs. Details on the activation functions and aggregation heuristics are available in Ji et al. (2020).
A set of multivariate time series from all visible active regions. Near-real time HARP series.
The probability of occurrence for the peak proton flux exceeding 10 pfu for proton energies >10 MeV over the next 24 hours and a binary all-clear flag.
- Heliosphere / Inner Heliosphere
Space Weather Impacts
- Solar energetic particles - SEPs (human exploration, aviation safety, aerospace assets functionality)
- Solar Energetic Particles
- Ji, A., Arya, A., Kempton, D., Angryk, R., Georgoulis, M. K., & Aydin, B. (2021, December). A modular approach to building solar energetic particle event forecasting systems. In 2021 IEEE Third International Conference on Cognitive Machine Intelligence (CogMI) (pp. 106-115). IEEE.
Code Languages: Python
Public Repository: https://bitbucket.org/gsudmlab/nrt-sep-all-clear-aggregator/src/V0.1.0/
- Anli Ji, Georgia State University (Model Developer)
- Berkay Aydin, Georgia State University (Model Developer)
- Dustin Kempton, Georgia State University (Model Developer)
- M Leila Mays, NASA GSFC CCMC (CCMC Model Host)
In addition to any model-specific policy, please refer to the General Publication Policy.