Speaker
Description
The Van Allen Radiation belts are highly dynamic in both strength and location, meaning that the belts are difficult to predict for spacecraft operators. Forecasting models exist, in part, to minimise any additional damage caused by this natural hazard. Both physics-based and machine learning models already exist; physics-based models allow for a deeper understanding of the system, and machine learning models offer a computationally cheap way to make a forecast but do not necessarily provide physical insight.
We present a collection of machine learning models capable of predicting if the Outer Radiation Belt crosses set percentile thresholds with considerable skill up to 3-days in advance, and some skill up to 6-days in advance. We use a Random Forest classification model to predict if the daily ~2MeV electron flux level across the Outer Radiation Belt exceeds thresholds from the 60th to the 95th percentiles. Each model shows a high level of accuracy at nowcasting and skill at forecasting up to 6 days in advance, a longer forecast than current operational models. Using feature importance, we determine the key inputs into each model in order to gain an insight into which drivers are important in driving increasing flux levels and over what timescales they have an impact. Crucially, we find that only a small number of geomagnetic indices are required to be able to forecast radiation belt fluxes with good skill, meaning that models such as these could be operationally viable for space weather stakeholders.
| Do you plan to attend in-person or online? | In-person |
|---|
