Prediction Of Solar Activity Based On Extreme Learning Machine And Echo State Network

Solar activity affects and destructs the global environment, so it is extremely important for its fast and accurate prediction and it has been widely concerned. Solar flares and sunspots are the main modes of solar activity. Based on the observation data of solar photospheric magnetic fields and sunspots, this paper establishes prediction model respectively with extreme learning machine and echo state network in order to achieve the prediction of solar flares and sunspots. Two main areas work of this paper are as follows:(1)A solar flare prediction method with an ordinal logistic regression model and a combination of several extreme learning machine models is proposed. Most of the existing solar flare prediction methods have low accuracy of the high-class flare forecasting. Although the algorithm from a combination of several BP neural networks has higher accuracy of the high-class forecasting, its training time is too long. Then the algorithm from a combination of several extreme learning machine models is proposed to forecast solar flares, this algorithm doesn’t need multiple iterations and has fast learning speed. Therefore, the solar flare prediction model from a combination of several extreme learning machines has a shorter training time. In order to further improve the prediction accuracy of the high-class flare, an ordinal logistic regression model is introduced to the flare prediction. Firstly, the ordinal logistic regression model is used to map three magnetic parameters (i.e. the total unsigned magnetic flux, length of the strong-gradient magnetic polarity inversion line and total magnetic energy dissipation) into four probabilities. Secondly, the model from a combination of several extreme learning machines is utilized to forecast the solar flare. Finally, the cross-validation approach is applied to test the experiments. The results of the experiments show that the proposed method not only has fast learning speed, but also has higher accuracy of high-class flare prediction.(2)A higher accuracy sunspot prediction model of wavelet transform and echo state network is put forward. Currently the traditional neural networks for sunspot numbers and sunspot areas prediction have the feature of high computational complexity, slow convergence and lack of memory. Therefore an echo state network with memory capacity is introduced into the sunspots prediction, the results of the sunspot numbers and sunspot areas prediction show that the method based on echo state networks has higher accuracy than the methods based on BP neural network and RBF neural network. Further, we introduce the wavelet transform into the sunspots prediction, presenting a prediction model of wavelet decomposition and echo state network. The original time series can be decomposed into several time series with wavelet decomposition, and then we make the several time series by signal reconstruction technique. On the one hand, the stability of the time series after signal reconstruction is better than the original time series, so it can improve the prediction accuracy. On the other hand, the low frequency and high frequency after signal reconstruction are put into the echo state network as the input unit of the reservoir, the high frequency in a certain sense can overcome the ill-posed problem which in the traditional echo state network models. Experiments with sunspot numbers and sunspot areas show that the proposed method has better predictive precision according to different training steps and prediction steps.