• Team Members: Ghanashyam Khanal, Nabin Malakar, Shovit Bhari and Shree K Bhattrai.

The goal of this challenge is to develop models for forecasting Dst (Disturbance Storm-Time Index) that 1) push the boundary of predictive performance 2) under operationally viable constraints 3) using specified real-time solar-wind data feeds.

• Feature Engineering:

  • Using our knowledge from Geo physics we dropped a few features which were very unlikely to affect the final prediction and picked only a few features for the LSTM model.

• Data amputation:

  • Filled the missing data in smoothed_ssn using forward fill and solar_wind using interpolation
  • Converted the data to hourly resolution picking both mean as well as the standard deviation.

• Feature Scaling:

  • Used Standard Scaler to scale the data. This turned out to be pretty important for accuracy score.

• Tried several things
  • play with batch_size
  • with or without the dropout (with dropout gave better RMSE)
  • Add batch normalization, it didn't improve the result
  • with or without activation function - tried Relu, sigmoid. Relu didn't improve the results

Following is a screenshot of one of the model architecture.

Following is the loss plot for train data (blue curve) and validation data (ornage curve) for one of our best models.

Our final performance was inside top 50 (top 8%) submitted by Shree Bhattrai.

Here is a quick fact about total participants.

• Benchmark submission
• DrivenData Git
• DrivenData Website
• https://machinelearningmastery.com/how-to-develop-lstm-models-for-time-series-forecasting/
• blog on multivariate time series forecasting
• https://github.com/Pythonyatra/MagNet
• VAR Tutorial
• Nice tutorials on Time-Series Analysis