Skip to Main content Skip to Navigation

Machine Learning in Space Weather

Mandar Chandorkar 1, 2
2 TAU - TAckling the Underspecified
LRI - Laboratoire de Recherche en Informatique, Inria Saclay - Ile de France
Abstract : The study of variations in the space environment between the Sun and the Earth constitutes the core of space weather research. Plasma ejected by the Sun couples with the Earth’s magnetic field in complex ways that determine the state of the Earth’s magnetosphere. Adverse effects from space weather can impact communication networks, power grids and lo- gistics infrastructure, all crucial pillars of a civilization that is reliant on technology. It is important to use data sources, scientific knowledge and statistical techniques to create space weather forecasting and monitoring systems of the future. This thesis aims to be a step towards that goal. The work is organised into the following chapters. In chapters 4 and 5, we develop probabilistic forecasting models for predicting geo-magnetic time series. Combining ground based and satellite measurements, we propose a gaussian process model for forecasting of the Dst time series one hour ahead. We augment this model with a long short- term memory (LSTM) network and produce six-hour-ahead probabilistic forecasts for Dst. Quantifying uncertainties in the dynamics of the Earth’s radiation belt is an important step for producing ensembles of high fidelity simulations of the magnetosphere. In chapter 6, we infer uncertainties in magnetospheric parameters, using data from probes orbiting in the radiation belts, by com- bining simplified physical models of the radiation belt with Markov Chain Monte Carlo techniques. Machine Learning in Space Weather vIn time-varying systems, it is often the case that cause and effect don’t occur at the same time. A prominent example of this time-lagged behaviour is the Sun-Earth system. Particles ejected from the Sun, also called the solar wind, reach the Earth’s magnetosphere after a time delay which is dynamic and uncertain. In chapter 7, we propose a novel neural network based method, called Dynamic Time Lag Regression (DTLR), for predict- ing time-lagged effects of events. We apply the DTLR methodology to the problem of near-Earth solar wind forecasting from heliospheric data.
Complete list of metadata

Cited literature [214 references]  Display  Hide  Download
Contributor : Cyril Furtlehner <>
Submitted on : Tuesday, January 7, 2020 - 3:21:12 PM
Last modification on : Sunday, May 2, 2021 - 3:30:33 AM


Files produced by the author(s)


  • HAL Id : tel-02430788, version 1


Mandar Chandorkar. Machine Learning in Space Weather. Machine Learning [cs.LG]. Université of Eindhoven, 2019. English. ⟨tel-02430788⟩



Record views


Files downloads