Answers are grounded in this paper and cite their sources.
The recently proposed SA-ConvLSTM is a novel TEC spatiotemporal model which is effective in capturing both local spatial correlations and long-distance spatial correlations. However, it has two key limitations: firstly, the computation process of self-attention in SA-ConvLSTM consumes a large amount of GPU memory usage; Secondly, its spatial feature extraction relies on standard convolution operations and can only operate within a fixed geometric receptive field, resulting in insufficient adaptability to irregular TEC changes. To address these issues, this paper made two improvements on SA ConvLSTM: (1) Transform self-attention computation to frequency-domain implementation, reducing GPU memory usage; (2) Replace standard convolution with deformable convolution, allowing adaptive spatial feature extraction through learnable offsets. Based on these two improvements, we proposed a novel recurrent unit named FSA-DConvLSTM. Then, we developed a TEC spatiotemporal prediction model ED-FSA DConvLSTM by organizing FSA-DConvLSTM with an encoder decoder structure. Finally, we verified the effectiveness of our improvement by experiments. The ablation experiment reveals that calculating attention within the frequency-domain has reduced the memory usage by over 70%. And the deformable convolution reduced theRMSEby 3.15%, 2.76%, and 2.77% in 2015, 2016, and 2019, respectively. In addition, based on comparative experiments with five state-of-the-art TEC spatiotemporal prediction models (IRI 2020, C1PG, ConvLSTM, ConvGRU, and PredRNN), the ED-FSA-DConvLSTM demonstrates superior performance. Furthermore, the FSA DConvLSTM is generalizable and can be adapted as a basic block for other spatiotemporal prediction tasks. Our method of reducing memory usage can provide reference for lightweighting other deep learning models.
No takes yet. Share an insight, caveat, or question.
Xiao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: