GLSP‐EDConvLSTM: Enhancing TEC Prediction Performance by Simultaneously Perceiving Local and Global Spatial Features

Abstract Total Electron Content (TEC) is influenced by local and global scale factors: localized effects causing fine‐grained differences in small areas, and global‐scale effects driving synchronized variations across all over the world. This underscores the necessity of modeling both global and local features for high‐performance TEC prediction. The existing deep learning models used for TEC prediction mainly rely on two‐dimensional convolution, which is effective in extracting local spatial features but not effective enough in capturing global spatial dependencies. To simultaneously model local spatial dependencies and global spatial dependencies in TEC, we first developed a novel global and local spatial perception (GLSP) module by integrating Vision Mamba for global spatial perception, two‐dimensional convolutions for local spatial perception, and a skip connection for preserving the original data. The GLSP was then integrated into ConvLSTM to form GLSP‐EDConvLSTM, where GLSP was utilized to pre‐extract global and local spatial features, reducing the burden on ConvLSTM's spatial feature extraction, allowing it to focus on temporal dynamics, and overcoming the deficiency of ConvLSTM due to its fixed local receptive field, which lacks global spatial dependency perception. To validate the proposed model, we tested it on a 10‐year GIM data set against 5 mainstream TEC methods (C1PG, EDConvGRU, EDConvLSTM, EDPredRNN, VMRNN). Results have shown that the proposed model achieved reductions of 23.7%, 5.8%, 4.5%, 4.0%, and 5.5% in 2019 and 21.0%, 5.3%, 3.0%, 4.8%, and 7.8% in 2024, along with excellent accuracy across latitudinal bands and reliability during extreme geomagnetic storms.