DstLive is a model for real-time predictions of the Dst Geomagnetic index. The 'ground-truth' for Dst is assumed to be the one issued by the World Data Center for Geomagnetism, Kyoto (https://wdc.kugi.kyoto-u.ac.jp/dstdir/).
DstLive predicts Dst from 1 to 6 hours ahead, with associated probabilities. The model is based on multi-fidelity boosted neural networks (Hu, Camporeale, Swiger, 2022), that ingest real-time data solar wind observed by ACE at the 1st Lagrangian point (https://www.swpc.noaa.gov/products/ace-real-time-solar-wind). In addition, the time history of Dst is used.
The model requires the latest observed Dst to produce a forecast. However, that value is not immediately available at the top of an hour. Hence, the model runs every 15, 30, 45 minutes past the hour, checking if the Dst for that hour has been released from Kyoto. However, in the current version, forecasts are always issued for top of the hour times, regardless of when they where issued (hence, they are effectively less than 6 hours ahead). This feature will be improved in future releases. Also, notice that it might occur that a given Dst value (so-called quicklook Dst) can later be changed, even within the same hour. In that circumstance, one will notice a discrepancy in the top panel between the last value shown as a blue line (observed Dst) and the first value shown as a black line (predicted Dst). Those differences can be tracked in the timeseries labeled Observed Dst at time of model run (see below).
The following inputs are ingested by the neural networks: solar wind density, velocity, amplitude of Interplanetary Magnetic Field (IMF), z-component of IMF, IMF clock angle, Earth's dipole tilt angle, and Dst. Solar wind quantities and IMF are taken in real-time from https://services.swpc.noaa.gov/text/ and the Dst index from the World Data Center for Geomagnetism, Kyoto (https://wdc.kugi.kyoto-u.ac.jp/dstdir/). See the reference paper (Hu, Camporeale, Swiger, 2022) for details.
The model outputs predictions in terms of Gaussian distributions, where the DstLive prediction, shown as a continuous line is the mean of the Gaussian distribution, and the shaded area is one standard deviation. In other words, the model predicts a 68% probability for Dst to lie within the shaded region. In the default settings, the top panel shows the observed Dst for the past 24 hours, and the predicted Dst for the future 6 hours (prediction are always made for the top of the hour), each shown with its confidence interval (one standard deviation). However, the user has the possibility to overlay predictions made with different time lags (t+1, t+2,... t+6 being 1, 2,...,6 hours ahead). The timeseries labeled Observed Dst at time of model run shows the Dst used as inputs and available at the time when the predictions were issued. Sometimes those values are adjusted at later times. Finally, the Dst predicted from the U. Michigan Geospace model, run operationally at the NOAA Space Weather Prediction Center (https://www.swpc.noaa.gov/products/geospace-geomagnetic-activity-plot) can be shown, labeled as Michigan Geospace V2 model results. Note that that is not the official SWPC forecast, instead it is a model guidance which forecasters can use to assist them in generating the official forecast. Real-time outputs are obtained from https://nomads.ncep.noaa.gov/pub/data/nccf/com/swmf/prod/, while historical storms have been simulated offline by Qusai Al Shidi at the University of Michigan (those results have been reported in Al Shidi et al., 2022).
A. Hu, E. Camporeale, B. Swiger (2022) Multi-Hour Ahead Dst Index Prediction Using Multi-Fidelity Boosted Neural Networks under review https://arxiv.org/abs/2209.12571 Al Shidi, Q., Pulkkinen, T., Toth, G., Brenner, A., Zou, S., & Gjerloev, J. (2022). A Large Simulation Set of Geomagnetic Storms–Can Simulations Predict Ground Magnetometer Station Observations of Magnetic Field Perturbations?, Space Weather, https://doi.org/10.1029/2022SW003049
Please contact Enrico Camporeale (enrico.camporeale@noaa.gov) for any questions.