Residual Attention-BiConvLSTM: A new global ionospheric TEC map prediction model

The prediction of total ionospheric electron content (TEC) is of great significance for improving the accuracy of global satellite navigation systems (GNSS). The existing TEC map prediction models are mainly structured by sequentially stacking spatiotemporal feature extraction units, which will lose fine-grained spatial features of TEC maps due to the sequential stacking of multiple convolutional layers, resulting in insufficient model accuracy; It may also cause gradient vanishing or exploding problems due to multi-layer stacking. Inspired by the idea of residual attention network, we add a residual attention module to the TEC map prediction model, proposing the Residual Attention-BiConvLSTM model. The residual attention module in our model can simultaneously extract coarse and fine-grained spatial features and weight them. This article conducted comparative experiments with ConvLSTM, ConvGRU, ED-ConvLSTM, and C1PG on global TEC maps. The experimental results showed that the RMSE, MAE, MAPE and R-2 of our Residual Attention-BiConvLSTM are superior to the comparison models in both high and low solar activity years. This article also compared the predictive performance of five models in a magnetic storm event, and the experimental results show that during a large magnetic storm, the model proposed in this paper is similar to C1PG and superior to the other three comparative models. The research work of this article provides a new approach for building ionospheric map prediction models.