Presentation Information
[PEM04-04]Solar Flare Prediction Studies Using Universal Time Series Predictor UFCORIN
*Takayuki Muranushi1, Yuko Hada Muranushi2, Takuya Shibayama4, Hiroaki Isobe2, Shigeru Nemoto3,2, Kenji Komazaki3, Kazunari Shibata2 (1.RIKEN Advanced Institute for Computational Science, 2.Kyoto University, 3.BroadBand Tower, Inc., 4.Nagoya University)
Keywords:
Space Weather Forecast,Flare Forecast
We have been studying space weather forecast using time-series prediction engine UFCORIN(Universal Forecast Constructor by Optimized Regression of INputs.) In our studies (Muranushi et al. 2015), we have compared 6'160 different prediction strategies that uses subset of wavelet features of SDO/HMI images as well as GOES past light curves.
Use of TSS (True Skill Statistics) as the indicator of flare forecast performance has been widespread since it is proposed by Bloomfield et. al(2012). However, we found that variation of bare TSS values over different cross-validation (CV) data sets is too large, so that we cannot measure significant difference between different forecast strategies. We found by using the $z$-value, or standard deviation of TSS, we can distinguish such strategies that show forecast performance consistently better than the average. We suggest the use of $z$-value as a method of finding good forecast strategies from thousands of candidates.
In our studies, the laregest TSS for X,M, and C class flare forecast, were $0.75\pm0.07$, $0.48\pm0.02$, and $0.56\pm0.04$, respectively.
Based on (Muranushi et al. 2015), we have been operating real-time flare forecast server since August 2015. The system have been making forecast every 12 minutes, except for some down times. We would also like to report on the latest state of this experience.
Use of TSS (True Skill Statistics) as the indicator of flare forecast performance has been widespread since it is proposed by Bloomfield et. al(2012). However, we found that variation of bare TSS values over different cross-validation (CV) data sets is too large, so that we cannot measure significant difference between different forecast strategies. We found by using the $z$-value, or standard deviation of TSS, we can distinguish such strategies that show forecast performance consistently better than the average. We suggest the use of $z$-value as a method of finding good forecast strategies from thousands of candidates.
In our studies, the laregest TSS for X,M, and C class flare forecast, were $0.75\pm0.07$, $0.48\pm0.02$, and $0.56\pm0.04$, respectively.
Based on (Muranushi et al. 2015), we have been operating real-time flare forecast server since August 2015. The system have been making forecast every 12 minutes, except for some down times. We would also like to report on the latest state of this experience.