Prediction of thermospheric temperature over the South Pole based on two-layer LSTM network

This study presents a new two-layer LSTM network-based model, which improves the accuracy of thermospheric temperature over the South Pole simulated by MSIS2.0 model. A dataset is constructed using temperature data measured by the South Pole FPI from 2000 to 2011 along with corresponding temperature derived from MSIS2.0 model, F10.7 and Ap indices, which are the input parameters of the first LSTM network layer. The first LSTM layer combines these inputs into a one-dimensional time series, while the second LSTM layer extracts temporal features from the output of the first layer. The proposed LSTM-based model shows better performance in predicting FPI observations compared to the empirical MSIS2.0 model during both geomagnetically quiet and disturbed periods. For the year 2011, the mean absolute error between the MSIS2.0 model and FPI data is 53.460 K, whereas the LSTM model reduces it to 34.024 K. The euclidean distance analysis also demonstrates better performance of the LSTM model. This study illustrates the potential of applying a two-layer LSTM network to optimize model simulations in upper atmosphere research.