CGConvLSTM: A Spatial‐Aware Ionospheric Total Electron Content Spatiotemporal Prediction Model

Abstract

Abstract The current deep learning models used in ionospheric total electron content (TEC) spatiotemporal prediction models rely on standard convolutions to extract spatial features. The weight sharing mechanism in standard convolution are position independent, which has the disadvantage of low spatial perception efficiency and ineffective in representing complex TEC spatial variations. CoordGate is a novel solution that adopts a multiplicative gate and a coordinate encoding network to enable spatially‐aware convolution. In this paper, we use CoordGate to extend ConvLSTM and propose a new spatiotemporal feature extraction unit CGConvLSTM. Then we construct a TEC spatiotemporal prediction model named ED‐CGConvLSTM based on CGConvLSTM. To verify the proposed ED‐CGConvLSTM, this paper compared it with 5 mainstream models normally used in TEC prediction, such as C1PG, ConvLSTM, ConvGRU, PredRNN, and IRI2020, on a data set containing 6 years of Global Ionospheric Maps (GIMs), with 4 years for the training set and 2 years for the test set. Experimental results show that compared to IRI2020, C1PG, ConvLSTM, ConvGRU, and PredRNN, ED‐CGConvLSTM's RMSE decreased by 60.97%, 17.97%, 10.80%, 13.68%, 7.96% in high solar activity years, and by 58.33%, 19.75%, 12.16%, 16.67%, and 7.80% in low solar activity years. Moreover, the superiority of ED‐CGConvLSTM was further verified from both spatial and temporal perspectives, as well as in extreme situations.