基于深度学习的太阳10.7 cm射电流量日值的中期预报

针对太阳10.7cm射电流量中期日值预报问题,采用深度学习方法,建立了一个典型的基于多层感知器模型的神经网络。该网络采用1个包含90个神经元的隐含层,实现了一种非参数的时间序列自回归模型。预报中不仅考虑历史日值,还考虑了历史预报误差。模型根据前27d的历史数据实现了未来27d的日值预报。通过对50多年数据的训练和试验分析,该方法在短期和中期预报上较传统方法的相对误差明显降低。特别是模型经一次训练后,参数可以完全固定,不同于以往研究参数需要每天滚动更新,大大简化了日常预报,同时极为有利于模型在其他相关应用中的推广。

Deep Learning for Mid-Term Forecast of Daily Index of Solar 10.7 cm Radio Flux

For mid-term forecast of the daily index of solar 10.7 cm radio flux with deep learning method,a neural network based on classical multi-layer perception model is proposed.The network contains only one hidden layer with 90 neutrons,and an autoregressive model of time series is implemented non-parametrically.In the forecast,historical daily indices as well as historical forecast error are considered.The model gives forecast of next 27 days with values of past 27 days.The network is trained and validated with historical data over 50 years,and the result clearly shows that the mean relative error is significantly reduced compared to the traditional methods.Unlike most of previous studies,in which the parameters of the model need to be rolling-updated,the parameters are fixed after the training with this model.The proposed model greatly simplifies daily operation of forecast and is extremely advantageous to the promotion in other applications.