Research On The Sunspot Activity Prediction Method Of Two-way GRU Network Based On Attention Mechanism

Solar activity is closely related to human life,at present,the sun is in the prime of its life cycle,and solar activity is more intense.Sunspot is the most common solar activity phenomenon on the solar surface,and it is also the most direct and efficient indicator of solar activity.The increase of sunspot number is often accompanied by the increase of solar burst activity,which leads to disastrous space weather.The disastrous space weather will cause the disturbance of the earth's magnetic field,interfere with radar and radio communication,and indirectly affect the earth's climate change.Carrying out sunspot prediction can help humans prevent disasters.In the traditional prediction methods,some methods rely on the empirical correlation model to predict,making the prediction results generally delayed.Some methods use complex mathematical formula to do calculation,and the small adjustment of parameters will make the prediction result change greatly.The prediction results obtained by these methods are not as good as expected.In this regard,this thesis uses the "memory" feature of the recurrent neural network model of deep learning,adopts a bidirectional gate recurrent unit(Bi-GRU)network with attention mechanism,and combines batch prediction methods to predict sunspots number in the next ten years.,and use the predicted results as the basis for studying the 25 th solar activity cycle.In the design process of the experiment,we set the value strategy of the step size in the model according to the characteristics of batch forecasting,and put forward a method of building multivariate input data.The results prove that the model does achieve a good prediction effect,it not only can accurately predict the maximum amplitude of the cycle,the start value and end value of the cycle,but also ensure their accuracy at the time node.Through the evaluation and analysis of the model,we prove that the attention module plays an important role in the whole prediction model.Compared with the results of the one-way Long short-term memory no-attention mechanism model,we found that the Bi-GRU model with the attention mechanism performs more smoothly in the process of predicting,the change of the sunspot without excessive jitter.Moreover,the distribution of data in time is relatively uniform,and there is no obvious aggregation phenomenon.The model has a stronger ability to learn the detailed features of the sunspot data,and can reflect the double-peak characteristics of the sunspots in some prediction results.Finally,we use the Bi-GRU network with attention mechanism to predict the number of sunspots in the next decade,providing theoretical basis and scientific prediction for the analysis of the 25 th solar cycle.The results show that the sunspots maximum amplitude of the 25 th solar cycle will reach 163.8 and appearing in September 2023,solar activity will be more intense than the 24 th cycle.