Dst Forecasting#
The disturbance storm-time index, Dst, measures the depression of the horizontal geomagnetic field at low latitudes, produced mainly by the storm-time ring current. It is a standard summary of geomagnetic storm strength, and forecasting it from the solar wind is a classic space-weather problem.
The notebooks forecast hourly Dst one hour ahead from OMNI2 data for 2010-2015. Each input holds the previous three hourly values of solar-wind speed \(V\), the GSM \(B_z\) component, field magnitude \(|B|\), and Dst. The network is
12 inputs → 50 ReLU → 30 ReLU → 1 output
The years are split in time: 2010-2013 for training, 2014 for validation, and 2015 for testing. Windows never cross a data gap or a split boundary, and the input scaling is fitted on the training years only.
The baseline is persistence, which assumes Dst does not change:
Because Dst varies slowly outside storms, persistence is hard to beat at a one-hour horizon. The notebooks report MAE, RMSE, and \(R^2\) for all hours and for storm hours separately, the MSE skill score relative to persistence, and a close look at the St Patrick’s Day storm of 17 March 2015.