摘要：The feature-based Bayesian method previously developed by Chen et al. (1996, 1997) to predict the occurrence, severity, and duration of large geomagnetic storms has been run on a daily basis on the Wind/Magnetic Fields Investigation (MFI) data from January 1996 until March 2010. The algorithm uses as input real-time solar wind magnetic field data obtained at the L1 Lagrange point, and the output is the probability prediction of the magnetic field structure of the upstream solar wind that has yet to arrive, and its geoeffectiveness, where geoeffectiveness is measured by the traditional Dst index. The performance characteristics of the method are evaluated using a four-level contingency table: nonstorm disturbances (−80 nT Dst ≤ − 50 nT), weak storms (−120 nT Dst ≤ − 80 nT), moderate storms (−160 nT Dst ≤ − 120 nT), and strong storms (Dst ≤ − 160 nT). It is found that the greater the level of disturbances, the more accurate the prediction is. With moderate and strong storms combined, the algorithm correctly predicted 30 out of 37 storms (81%). The false negatives are caused by solar wind structures with short durations (≲ 1–2 hrs) of strong southward magnetic field (say, Bz ≲ −30 nT), which are sparsely represented in the probability distribution functions constructed using prior solar wind data (OMNI data set from 1973–1981). The algorithm does not predict the storm onset time, but the results of the present and previous tests show that the average warning time ranges from a few hours to a maximum of 10–15 hours.
