Exploring Heuristics in Full-Disk Aggregation from Individual Active Region Prediction of Solar Flares
Abstract
Ensemble modeling can boost the predictive confidence of base learners (weak learners) by producing a more optimal model suitable for operational forecasting. In cases where the base learners have uniformity in the problem formulation and datasets, it becomes a relatively easier task to generate a new meta-model that enhances prediction capabilities. In this work, we deal with coupling two solar flare prediction models in which our base learners are trained with two different paradigms: (i) a full-disk mode, where prediction models are trained using full-disk line-of-sight magnetograms and deep learning architectures to issue a full-disk flare probability, and (ii) an active region-based mode, where models utilize multivariate time series data and a multivariate Time Series Forest (TSF) classifier to issue a prediction for each valid active region individually. So far in the literature, individual active region forecasts are aggregated by computing the probability of flare from at least one active region assuming conditional independence and then these flare probabilities are used to compute a full-disk flare occurrence probability by means of a meta-model. However, our empirical analysis on active region aggregation shows that the method of aggregation can be considered rather subjective depending on the distribution of individual models' output. We observe that exploring different heuristics while aggregation impacts the predictive performance of the AR-based models and consequently the meta-models. This shows a need for searching an AR-aggregation heuristic that suits the overall probability distribution of our two base learners. In this research, we present different AR-aggregation heuristics in consideration to the flare probability distribution of our base learners and introduce a more effective heuristic for active region aggregation by verifying performance improvement using metrics such as the True Skill Statistic and the Heidke Skill Score.
- Publication:
-
44th COSPAR Scientific Assembly. Held 16-24 July
- Pub Date:
- July 2022
- Bibcode:
- 2022cosp...44.3457P