Three-hour predictions of the planetary magnetospheric Kp index are made using multilayer feed-forward neural network models. The input parameters for the networks are the Bz-component of the interplanetary magnetic field, the solar wind density n, and the solar wind velocity V, given as three-hour averages.

Every hour, the latest ACE solar wind data are downloaded from the NOAA's Space Environment Center in Boulder and then presented to the network model. Data processing results in the diagrams displayed above. The blue curves represent the solar wind data and the predicted Kp index. The predicted Kp index is compared with the estimated planetary K index from SEC (shown in red). The vertical line indicates the present time.

Real time Kp predictions from solar wind data using neural networks by F.Boberg, P. Wintoft and H. Lundstedt, published in Physics and Chemistry of Earth Vol 25 No 4 2000, can be downloaded here.

The accuracy of these Kp predictions is given by the diagram below. The root-mean-square error of measured minus predicted Kp is plotted as a function of measured Kp indicating an adequate accuracy for geomagnetic storm predictions.