Satellite clock bias prediction for BDS based on CNN-LSTM-Attention

Satellite clock bias constitute one of the key factors affecting the performance of the BeiDou Navigation Satellite System (BDS). However, existing post-processing precise satellite clock bias products fail to meet the demands of high-precision real-time positioning, navigation and timing (PNT) services. Therefore, to further enhance the precision and real-time capability of BDS applications, a combined satellite clock offbias prediction model integrating convolutional neural networks (CNN), long short-term memory (LSTM) networks, and Attention mechanisms is proposed. The proposed model first employs CNN to extract deep temporal features from time segments, then utilizes the LSTM model to establish longer-term dependencies. Subsequently, the Attention mechanism is introduced to adaptively weight the hidden layers of the LSTM, balancing the capture of global and local features. This ultimately achieves high-precision real-time forecasting of satellite clock offsets. Finally, the proposed model undergoes comprehensive forecasting performance comparisons against auto-regressive integrated moving average (ARIMA), grey model (GM), CNN, and LSTM models. Results demonstrate that for 24 h forecasting tasks, the proposed model achieves prediction accuracy improvements of 73.0%, 76.6%, 88.5%, and 74.8% respectively over ARIMA, GM, CNN, and LSTM models.