计算机科学
人工智能
传输(计算)
学习迁移
理论(学习稳定性)
工程类
工作(物理)
钥匙(锁)
领域(数学)
光学(聚焦)
作者
Jiapeng Liu,Kuo Chi,Ting Su,Yi Yang
出处
期刊:
日期:2026-04-24
卷期号:: 137-140
标识
DOI:10.1109/eibdct69742.2026.11566741
摘要
Accurate prediction of the Maximum Usable Frequency (MUF) is critical for shortwave sky-wave communication and adaptive frequency selection. Traditional empirical models lack temporal resolution, while data-driven deep learning methods suffer from overfitting due to measurement scarcity on specific links. This paper proposes a model-data dual-driven transfer learning network (CNN-BiGRU-MHA) incorporating multi-source space weather features. A six-dimensional physical feature system is constructed, and CNN-BiGRU with Multi-Head Attention (MHA) and Global Average Pooling is employed for spatiotemporal feature extraction. A two-stage strategy pre-trains on large-scale simulation data, then fine-tunes on limited maritime link measurements. Experiments on a low-latitude maritime link demonstrate that the proposed method significantly reduces prediction error during geomagnetic disturbances, achieving substantially higher accuracy than both the VOACAP empirical model and baseline deep learning networks under small-sample conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI