| ACE News Archives | ACE News #68 - Feb 20, 2003 |
ACE News Archives |
The locally accelerated component of an Interplanetary (IP) shock is referred to as an Energetic Storm Particle (ESP) event due to its strong association with geomagnetic storms. An ESP event can encompass ion activity for several hours upstream and downstream of the IP shock. The gradually increasing intensity precursors observed upstream of such shocks evolve in a characteristic manner that has been well documented in the literature, and can be used to predict the passage of an IP shock. During an ESP event, < 10 MeV ion intensities can increase by orders of magnitude. Because of the radiation hazards to astronauts and space-embedded technology posed by energetic particles, there is considerable interest in forecasting large IP shock-driven particle events.
In this work, we use historical EPAM data to train an ESP event forecasting algorithm. Our approach centers on the observation that large ESP events are often preceded by identifiable signatures in the low energy ion particle intensity (as shown in the left figure for the September 6, 2000 ESP event). Using these identifiable signatures as input, we train an artificial neural network to predict the time remaining until the maximum intensity of an ESP event. Training was performed using 37 of the 56 selected ESP events observed on ACE. The performance of the network was assessed by having it forecast arrival time count-downs for the 19 shocks not used during training. The figures above show the input sequence and forecast for 1 of the 19 shocks used in testing. The left panel shows time intensities of the EPAM ion channels from 46 keV to 1.9 MeV (from highest to lowest intensities, respectively). Labeled vertical lines indicate the onset time and the peak time. The onset time is selected by the algorithm automatically based on velocity dispersion and anisotropy of the particle data to trigger the prediction part of the algorithm. The prediction part of the algorithm only uses data after the onset as input.
The panel on the right shows a comparison of the forecast and actual time until maximum intensity. Since the true output is always a count-down until the intensity peak, the ideal output is always a straight line with a slope of -1. This test sequence exhibits very good performance. The neural network briefly under-estimates the shock arrival time by about ten hours when it is still 44 hours away, and it over-estimates the arrival time by about ten hours when it is 18 hours away, but overall its prediction aligns well with the true arrival time. Results over the entire test set are very encouraging; the average uncertainty in the prediction 24 hours in advance is 8.9 hours, while the uncertainty improves to 4.6 hours when the event is 12 hours away.
Contributed by George Ho, Jon Vandegriff and Kiri Wagstaff of Johns Hopkins University Applied Physics Laboratory. This work was partially supported by NASA LWS Grant NAG5-10636.
Last modified 20 February 2003, by