Abstract
The predictive learning of total electron content (TEC) spatiotemporal sequences aims to generate future TEC maps by learning from historical data, where both the spatial appearances and temporal variations are crucial for accurate predictions. However, the state-of-the-art TEC map prediction models typically employ sequential stacking of ConvLSTM, ConvGRU, and their variants. These models focus more on modeling temporal variations, and the spatial features extracted from the historical sequence are highly abstracted, resulting in the fine-grained spatial appearances not being adequately memorized or transmitted, leading to fuzzy prediction results during storm time. In this paper, we used PredRNN to propose a storm-time ionospheric TEC spatiotemporal prediction model with multichannel features, named Multichannel PredRNN, which can simultaneously remember the temporal patterns and spatial appearances in input sequence. The temporal memory as well as the spatial memory are updated repeatedly over time, ensuring that both temporal memory and spatiotemporal memory are fully utilized in prediction. According to Dst index, 60 magnetic storm events from 2011 to 2019 were selected as the dataset. We first discussed the impact of feature combinations on predictive performance. The results show that using multichannel feature (TEC + Dst&F10.7), the Multichannel PredRNN and the comparison models ConvGRU and ConvLSTM have the best prediction performance. Then we used the optimal feature combination for prediction. We compared Multichannel PredRNN with IRI-2016, COPG, ConvLSTM and ConvGRU under various conditions, including the entire test magnetic events, periods of quiet and storm, different phases of geomagnetic storms, and the most severe geomagnetic storms. Finally, we compared the performance of different output steps. The experimental results indicate that in all cases, Multichannel PredRNN with dual memory state and zigzag flow is superior to four compared models.
