Innovative Uncertainty-Quantification based Ensemble Machine Learning Method on Extreme Space Weather Events
Abstract
We present an innovative ensemble machine learning (ML) method for forecasting extreme space weather events. Ensemble technique is a powerful tool to improve the accuracy of the predictions, and in fact it has been widely used for physical model predictions. One of the major difficulties for implementing ensemble techniques on ML methods is that typically ML models do not provide the uncertainty associated with a prediction. Hence, an uncertainty-quantification (UQ) based Gated Recurrent Unit (GRU) method is developed, to forecast the uncertainty for the model predictions simultaneously.
We have implemented this method on two space weather applications, i.e., 1) a one-to-six-hour lead-time model that predicts the value of Disturbance storm time (Dst) using solar wind (SW) data; and 2) an geoelectric field model with 1-hour leading time using SW and SuperMag data. The first model can forecast Dst 6 hours ahead with a root-mean-square-error (RMSE) of 13.54 nT during more than 50 strong storm events in recent two solar cycles. This significantly outperforms physics based and prior empirical models. The E model 1-hr prediction also agrees well with ground truth data generated from SuperMag data and Magnetotellurics (MT) surveys.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNG43A..01H