LSTM神经网络在太阳F10.7射电流量中期预报中的应用

F_10.7指数作为大气密度经验模型的重要输入参量，其预报精度直接影响航天器轨道预报精度.研究发现，太阳活动表现出长时间尺度上平均11年和中短时间尺度平均27天的周期性变化特征.依据这一观测事实，基于长短期记忆单元（Long Short-term Memory，LSTM）递归神经网络方法进行F_10.7指数未来27天的中期预报.利用一个连续长时段F_10.7数据作为训练数据，构建LSTM神经网络训练和预测模型，分别预测太阳活动高低年未来27天的F_10.7指数.结果表明，太阳活动高年的第27天F_10.7指数预报平均相对误差最优可达10%以内，低年最优可达2%以内. 

Application of LSTM Neural Network in F10.7 Solar Radio Flux Mid-term Forecast

The F10.7 index is an important input parameter for the empirical models of atmospheric density,and its prediction accuracy directly affects the accuracy of spacecraft orbit prediction.The solar activity exhibited an average of 11 years on a long-term scale and a 27-day periodic variation on a short-term scale.Based on this observational fact,a 1 Long and Short Term Memory(LSTM)recurrent neural network method is proposed to conduct the mid-term forecast of F10.7 index for the next 27 days.Using a continuous long period of F10.7 data as training data,the LSTM neural network training is constructed,and the upper and lower bounds of model parameters based on empirical formula are determined.The method of trial and error is used to select the optimal model parameters,and the prediction models to predict solar activity of high and low years F10.7 index in the next 27 days are constructed.The results show that the average relative error of the 27 th day F10.7 index forecast for solar activity in the high year can reach about 10%,and can reach 2%or less in the low year.In 1998,the correlation coefficient between the predicted value of the F10.7 index on the 27 th day and the measured value was 0.60.