Spatiotemporal prediction of ionospheric total electron content based on SA-ConvLSTM

Existing models for spatiotemporal prediction of Total Electron Content (TEC) primarily rely on stacked ConvLSTM units and their variants. This TEC spatiotemporal prediction model based on ConvLSTM is effective in capturing local spatiotemporal dependence. However, due to the lack of units for storing long-distance spatial memory, the spatial dependencies of long-distance TEC are difficult to capture by ConvLSTM and its variants. To solve this problem, this paper proposes a spatiotemporal prediction model SA-ConvLSTM for ionospheric TEC based on the self-attention memory ConvLSTM. Based on ConvLSTM with short-term memory dependence, this model adds a self-attention memory (SAM) module with long-distance memory dependence. This enables SA-ConvLSTM to simultaneously consider both short-term and long-distance memory in TEC spatiotemporal prediction. In this paper, to verify the performance of SA-ConvLSTM, TEC grid data of three years of high solar activity and three years of low solar activity are selected in the region of 12.5°S-87.5°N, 25°E-180°E. We compare SA-ConvLSTM with the current state-of-the-art TEC spatiotemporal prediction models, including ConvGRU, ConvLSTM, PredRNN, Residual Attention-BiConvLSTM and the ionospheric prediction product C1PG provided by CODE. The results show that compared with the RMSE of C1PG, ConvGRU, ConvLSTM, PredRNN, and Residual Attention-BiConvLSTM the those of SA-ConvLSTM is reduced by 6.58%, 3.89%, 5.79%, 1.44% and 1.21% in high solar activity years; and by 13.42%, 10.26%, 11.40%, 3.20% and 4.37% in low solar activity years, respectively. In addition, the different months and latitudes regions of each model are compared. The results show that the prediction performance of SA-ConvLSTM is better in the vast majority of months and latitude regions. Finally, two magnetic storm events are selected to verify the prediction ability of SA-ConvLSTM in extreme cases. The results indicate that SA-ConvLSTM outperforms the comparison model in most phases of magnetic storms.