top of page

Predicting Storm Times using Machine Learning

Students from Cramlington Learning Village with Orbyts Fellows Kendra Gilmore and Vishal Singh

Times of strong geomagnetic activity (storm times) can cause a lot of harm to humans via the infrastructure that many societies have become somewhat reliant upon. For example, heating of the upper-layers of the atmosphere can impede long-range communications and the quality of GPS. The perturbations to Earth's magnetic field can induce electrical currents in large-scale conducting infrastructure, such as power lines and pipelines, leading to immediate damage or degradation. In space, satellites are at great risk of damage, particularly if not put into a special "safe mode", and astronauts are exposed to increased levels of radiation. To minimise such disruption, we therefore want to predict in advance when storm times could happen.

Using machine learning, here so-called "random forests", we investigated which physical parameters relating to storms are most important for their prediction. We applied our methods to the OMNI dataset (spacecraft measurements of solar wind and interplanetary magnetic field parameters that have been time-shifted to when they should be reaching Earth) and SuperMAG data (measurements of magnetic field perturbations from a network of ground-based magnetometers) to identify storm times. Using only the solar wind speed as an input, our model predicts only non-storms even when we have storms (see Figure 1 above).  On the other hand, the z-component of the interplanetary magnetic field (the component perpendicular to the Sun-Earth line) provides the best results from a singular input, showing that it is the most important parameter (see Figure 2 below). This is because when this z-component is orientated southwards, it can efficiently connect with Earth's magnetic field, allowing for the transfer of matter into the Earth system. We find that we can improve predictions by using all components of the interplanetary magnetic field, thus predicting both storm and non-storm times well (see Figure 3).



Figure 1: A confusion matrix used for evaluating the performance of our prediction model in which the only input was the solar wind speed, Vsw


Left: Figure 2

Right: Figure 3

bottom of page