Abstract The existing Total Electron Content (TEC) spatiotemporal prediction models are more suitable for extracting and utilizing high‐frequency features, but there are difficulties in dealing with low‐frequency features. Wavelet transform convolution (WTconv) is a new low‐frequency and high‐frequency feature extraction operation that has been proven to be effective in other fields. Inspired by WTconv, this paper proposed a novel spatiotemporal feature extraction unit WTConvLSTM. In addition, we proposed a channel and spatial attention (CSA) to focus on channel dependency of features and important spatial information. Based on WTConvLSTM and CSA we proposed a spatiotemporal prediction model CSA‐WTConvLSTM. This study first utilizes data from Solar Cycle 24 (1 January 2009 to 31 December 2019) to develop the model, assigning 2013 and 2017 to the validation set, 2015 and 2019 to the test set, and the remaining for training. On this data set, the proposed CSA‐WTConvLSTM was compared with an ionospheric empirical model (IRI2020), a TEC prediction product (C1PG) and three deep learning models (ConvGRU, ConvLSTM and PredRNN). The overall quantitative comparison has shown that, the of CSA‐WTConvLSTM decreased by 61.74%, 19.57%, 10.71%, 7.43%, 4.00% in 2015 and 60.24%, 23.37%, 13.31%, 8.40%, 3.59% in 2019 compared to IRI2020, C1PG, ConvGRU, ConvLSTM and PredRNN, respectively. We also conducted more detailed comparisons by month and different latitude regions, as well as comparisons during geomagnetic storm periods. Additionally, we used three independent years (1 January 2020 to 31 December 2022) from Solar Cycle 25 to test our model's effectiveness. All experimental results confirmed the effectiveness of our model.
Zhou et al. (Fri,) studied this question.