TEC Prediction Based on Att-CNN-BiLSTM

Abstract

Prediction of Total Electron Content (TEC) in the ionosphere is vital to improve the accuracy of satellite positioning, navigation and remote sensing systems. Most existing TEC prediction methods ignored the local variation patterns between various positions within the TEC sequence, resulting in limited prediction accuracy. To address this issue, this paper combined attention techniques, convolutional neural networks (CNN) and bidirectional long short-term memory networks (BiLSTM) to propose the Att-CNN-BiLSTM model. In the proposed model, CNN is used to extract positional features, BiLSTM is used to extract bidirectional temporal features, and attention technique is used to adaptively weight the features. This paper selected six locations in China, each with a six-year TEC sequence, including three years of high solar activity and three years of low solar activity. The paper first conducted ablation experiments, and the results showed that adding CNN and Attention can effectively improve prediction performance. Then, the proposed model was compared with LSTM and GRU. The experimental results show that compared with LSTM and GRU, the average RMSE of Att-CNN-BiLSTM in six regions decreased by 10.28% and 16.92% in high solar activity years, and by 11.82% and 8.92% in low solar activity years, respectively. The paper also conducted comparative experiments within one week of the magnetic storms, and the results showed that during the magnetic storm period, the RMSE of the proposed model decreased by 35.50% and 37.35% compared to LSTM and GRU. The R2 s of the proposed model are also higher than those of the comparison models in all cases.

Keywords

• Attention mechanism
• CNN-BiLSTM
• ionosphere
• total electron content
• local variation patterns