1. Kp Forecast Models S. Wing 1, Y. Zhang 1, and J. R. Johnson 2 1 Applied Physics Laboratory, The Johns Hopkins University 2 Princeton Plasma Physics Laboratory, Princeton University Questions: simon.wing@jhuapl.edu

2. 2 Outline 1.Background and motivations 2.Data Set 3.Method 4.APL Kp models 5.Magnetospheric predictability as a function of solar cycle 6.Comparison with previous Kp models 7.Summary

3. 3 Background and Motivations for developing Kp forecast models Moderate and high activities are notoriously difficult to predict [Joselyn, 1995]. Real-time magnetometer data can be used to calculate nowcast Kps, which could improve the accuracy of the forecast Kps. Why Kp? Kp is one of the most popular global indices. Kp has been playing significant roles in space weather, e.g., satellite drags etc. Many magnetospheric and ionospheric models require Kp as an input parameter, e.g., T89 magnetic field model, Fok ring current-radiation belt model, MSFM, OVATION, etc. The long uninterrupted Kp record since 1932 makes it ideal for studying solar-wind magnetosphere interactions, e.g., the solar cycle effects, etc.

4. 4 Data Set 1.Solar wind and IMF data spanning over 2 solar cycles : IMP8 (1975- 1999), Wind (1994-2000), and ACE (2000-2001). Data are publicly available from NASA CDAWeb. IMP-8 plasma data are publicly available from MIT IMP-8 website. 2.Historical/official Kps (1975-2001) are available at GFZ Postdam website.

5. 5 Method 1.The time granularity of the input and output of the model is 15 min. Solar wind/IMF were propagated to Earth and hourly averaged every 15 min. Kps are interpolated to 15 min resolution. 2.The solar wind, IMF, and Kp data set were randomly selected into 2 equal subsets: (1) training set and (2) testing set. 3.Neural Network (NN) has been used in space physics/weather, particularly as predictors or classifiers [e.g., Koons and Gorney, 1991; Vandegriff et al., 2004; Wu and Lundstedt, 1997; Newell et al., 1991; Wing et al., 2003]. 4.In this study, NN: Input: solar wind, IMF and nowcast Kp (in some models) Output: forecast Kp.

6. 6 APL Kp forecast models In order to satisfy different needs and operational constraints, three different models were developed: 1.APL model 1 Input: ACE solar wind/IMF parameters and nowcast Kp Output: ~1-hr ahead Kp forecast 2.APL model 2 Input: same as model 1 Output: ~4-hr ahead Kp forecast 3.APL model 3 Input: ACE solar wind/IMF parameters Output: ~1-hr ahead Kp forecast Wing et al., Radio Sci., 2003 Wing et al., JGR, 2005

7. 7 APL Kp forecast models 1.APL model 1 Input: ACE solar wind Vx, n, IMF |B|, Bz, and nowcast Kp Output: ~1-hr ahead Kp forecast Note: nowcast Kp algorithm has recently improved, achieving r > 0.9 [Takahashi et al., 2001].

8. 8 APL model 1 predicts low and high Kps very well The higher performance can be attributed to: 1.Larger data set. 2.Input nowcast Kp. APL Kp forecast models Evaluation based on data spanning over 2 solar cycles

9. 9 APL Kp forecast models The performance is worse (larger scatter) than APL model 1. It is still better than the previous models that provide 1 hr ahead forecast.

10. 10 APL Kp forecast models The performance is better than the previous models.

11. 11 Kp (magnetospheric) predictability as a function of solar cycle Papitashvilli et al. [2000] report that there is solar cycle variation in the average Kp. It would be interesting to determine if the accuracy of Kp (a proxy for the magnetospheric state) forecast based partly or entirely on solar wind/IMF has solar cycle dependence. Calculate the skill scores (TSS and GS [Detman and Joselyn, 1999]) for Kp forecast models for 2 solar cycles, 1975-2000. TSS: True Skill Statistics GS: Gilbert Skill

12. 12 Kp predictability as a function of solar cycle 0 = random forecast 1 = perfect forecast Costello NN Kp model predicts Kp more accurately near solar maximum than minimum.

13. 13 Kp predictability as a function of solar cycle The solar cycle dependence of APL model 3 While the scores are higher than those for Costello model, they still exhibit solar cycle variation, albeit with smaller amplitude. Training with a larger data set cannot eliminate the solar cycle effect completely. APL model 3 performs better during solar max than solar min.

14. 14 Kp predictability as a function of solar cycle The nonlinear response anti- correlates with sunspot number in every solar cycle since Kp record is kept.

15. 15 Comparison between Kp models Boberg et al. [2000] Operational at Lund Obs. APL model 1 (1 hour forecast) APL model 3, 1-hour forecast (purely driven by solar wind) APL model 2 (4 hour forecast) Costello NN Kp model [Costello, 1997] Input: IMF Bz, |B|, and |V| 1-hour Kp forecast NARMAX [Balikhin et al., GRL] Bz, |B|, |V|, N, 1-hour forecast

16. 16 The Kp models were developed at JHU/APL. They have been transitioned to USA NOAA SWPC and other space weather centers for operational use. For example, See http://www.swpc.noaa.gov/wingkp/

17. 17 Summary and Conclusion Extensive model evaluations based on data spanning over 2 solar cycles show that 1.Kp is more predictable during solar max than solar min. 2.Cumulant based information dynamics analysis of Kp shows that Kp (the magnetosphere state) is more strongly nonlinearly coupled with the past Kps during solar min than solar max [Johnson and Wing, 2004]. 3.(1) and (2) suggest that the magnetosphere is more externally driven during solar max (declining phase of solar max) than solar min, when internal dynamics play a more significant role.

18. 18