Research In Forecasting AE,AL Index Based On Neural Networks

Magnetosphere sub storm is the coupling result between solar wind and magnetosphere, the research of which is an important aspect in the solar-terrestrial physics. So forecasting magnetosphere sub storm is a very significant part of space environment prediction. In general, as AE, AL etc indices are used to inspect the turbulence in polar region when sub storms happen, they are the target indices in space environment prediction. It is advanced work to use mathematical tools such as nonlinear filter, neural network quantitatively forecasting sub storm indices.In this paper, we discussed the process of solar wind energy immerging in magnetosphere and springing sub storm. While the magnetic field turns south, it usually induces sustaining sub storm, indices sharply changing. The changeful characteristic parameters of solar wind and magnetic field have great influence on the changing of sub storm indices, which is theoretical base of calculating indices.Neural networks, as one of the most active direction in the research of artificial intelligence and machine learning, have powerful computing ability. Especially the BP (back propagation) neural network is the most widely model in application, having great value in data calculation and forecasting.We construct an all-joint BP neural network as the prediction tool, choosing new database, based on the demand of Chinese space environment prediction. The By, Bz components of magnetic field, solar wind velocity and solar wind proton density as input parameters, calculates AE and AL indices.The prediction show that our network model can forecast the trend of fluctuation of sub storm indices, has wonderful veracity in quantitatively computing indices value. The correlation between four input variables (By, Bz, v, n) and sub storm indices is very good. The best correlation coefficient even can be more than 86%.Meanwhile, we think that the durative condition of solar wind and magnetic field influences the later sub storm indices in one time point, considering the time length of input variables for 20, 40, 60min. With the time extends, the forecasting precision improves. There may be about 1 hour length of solar wind affecting the changing of indices. As the forecasting ahead time is 50-80min of our neural network model, it can be used in real prediction work.