DISCOVER-DSCOVR

Challenge Summary

How to develop geomagnetic activity forecast using the raw DSCOVR data directly as input?

Solution Overview

Data Preprocessing

DSCOVR and WIND data are from different sources with different properties. To solve this, we followed these DSP methods:

Forecasting Model

Transformer model to extend the data for few hours later after the existing/available data

Train Loss: 0.03
Validate loss: 0.06

Kp Prediction Model

We used bert-base-uncased, BERT model as a starter model and then we fine tuned it to serve our problem with less time

Our model is pre-trained on similar time series problem.

We used the encoder only (self-supervised) to extract features from data.

Then we trained our own decoder on 5 months of data.

The final loss is 0.8 on MSE.