Abstract
Global ionospheric total electron content (TEC) map prediction is important for improving the accuracy of global navigation satellite systems. There are two main issues with the current TEC prediction: (1) The deep learning models used for TEC prediction are mainly designed using a stacked structure. When stacking multiple layers, the input data will undergo continuous multi-layer convolution operations, leading to the loss of fine-grained features and the degradation of model performance; (2) The model optimization methods for TEC prediction are relatively outdated, mainly using manual optimization or grid search methods. To address these two issues, an automatic framework for global TEC map prediction and optimization is proposed, named as CGAOA-STRA-BiConvLSTM. It includes a global TEC map prediction model, STRA-BiConvLSTM, which can simultaneously extract both coarse-grained and fine-grained spatiotemporal features. It also contains an optimization algorithm, CGAOA, to optimize the model. We first experimentally verified the effectiveness of CGAOA. Then, the effectiveness of STRA-BiConvLSTM was verified through ablation experiments. Finally, we conducted comparative experiments from multiple perspectives between our framework and 5 mainstream methods: C1PG, C2PG, ConvLSTM, ConvGRU, and ED-ConvLSTM. The results show that in all cases, the proposed CGAOA-STRA-BiConvLSTM outperforms the comparative models.