利用混合模型LSTM-DNN进行全球电离层TEC map的中短期预报

近年来,基于深度学习技术的长短期记忆(long short-term memory,LSTM)网络相关预报算法在空间天气的预测方面得到广泛应用,但存在预测误差随时间堆叠的缺陷,因此只能进行有限的短期预测.为解决这一问题,文章将太阳风参数、太阳黑子数、地磁活动水平指数Ap以及磁暴环电流指数Dst作为预报因子加入模型,建立...

Short- and medium-term forecasting of global ionospheric TEC map based on hybrid model LSTM-DNN

The Long Short-Term Memory (LSTM) related prediction algorithm based on the deep learning technology is widely used in the space weather prediction in recent years. Due to its shortcoming of stacked prediction errors against time, the LSTM-related prediction algorithms can only be used for limited short-term predictions. To improve its performance in the model, we add the solar wind parameters, the sunspot number, the Ap index representing the geomagnetic activity level, and the Dst index corresponding to the magnetic storm ring current as the predictors, to establish a hybrid model based on the LSTM and the DNN (deep neural networks) for medium and short-term forecasting of the global ionosphere TEC map, which can significantly inhibit the increasing trend of the forecast error over time. Test results show that compared with the LSTM model alone, the LSTM-DNN hybrid model enjoys a similar accuracy for the 24-hour prediction of the ionosphere. The average relative accuracy of the 48-hour ionosphere forecast is increased from 79.30% to 81.18%, and that for the 144-hour ionospheric prediction is increased from 64.97% to 77.64%.