Non-linear predictions of Ap by activity class and numerical value

Two neural network algorithms are applied to the short-term, 1 to 3 days, prediction of the Ap geomagnetic index. A multi-layer, back-propagation (MBP) network is used to implement a self-prediction filter for Ap and this provides a forecast of the numerical value of the index. A probabilistic neural network (PNN) is used to estimate the probability distribution of the Ap index, in six activity classes, and to provide a forecast of the single most likely activity class for each day. Both Ap and an index of solar activity, based on the daily reports issued by the Space Environment Services Centre (Boulder), are input to the probabilistic net. It is found that the numerical forecasts of the MBP filter are most accurate at low, non-storm, levels of activity. This non-linear method provides quantitatively better estimates of activity than are produced by an existing linear prediction filter, particularly with increasing forward forecasting lag. At high levels of the solar activity index the PNN is found to anticipate storm class Ap with around 60% accuracy in 1992 and 1993. Some derails of the algorithms and implementation issues are described. It is concluded that interplanetary field and solar wind data will be significant components of any of the possible future developments which are discussed.