基于深度学习的多步预测方法在太阳风速度预测上的应用

为对近地环境太阳风状况进行可靠预测,引入基于深度学习的多步预测方法来预测在太阳与地球之间的拉格朗日点1(L1)处距离输入观测数据序列未来24、48、72、96 h的太阳风速度.采用SDO的图像数据提取冕洞面积等特征信息,并从NASA OMNIWEB数据集提取其他输入特征,形成多变量的时序数据作为太阳风速度预测的输入信息...

Application of deep learning based multi-step prediction to predict solar wind velocity

In order to reliably predict the solar wind in the near earth environment, the deep learning multi-step prediction method is adopted to predict the solar wind speed at the Lagrange point 1 (L1) between the sun and the Earth in the next 24, 48, 72, and 96 hours. The SDO image data are used to extract the area of the coronal hole and other input features from the NASA OMNIWEB data set to form the multivariable time series as the input information for the solar wind speed prediction. In the multi-step prediction method, the output is no longer the solar wind speed at a certain moment in the future, but the solar wind speed for several hours around that moment at the same time. Three deep learning prediction models are used respectively: the single LSTM model, the encoder-decoder LSTM model and the CNN-LSTM encoder-decoder model as the base model, and the multi-step output data are included for the comparative analysis. Experiments show that, compared with the single-step output prediction, the multi-step prediction improves the correlation coefficient and the forecast accuracy of the solar wind speed forecasts in the next 24, 48, 72, and 96 hours. For the 24-hour prediction, the best experimental results are achieved: the correlation coefficient with the observed data is 0.789, the RMSE is 62.469 km/s, and the MAE is 46.476 km/s.